# Feature × Provider matrix — OpenAI · Anthropic · xAI · Gemini
**Status:** synthesis of `generated/*.json` (endpoints, parameters, tools, models, headers, pricing, streaming-events, webhook-events, deprecations) and the domain pages under `docs/`; every status shown is the status recorded in those files (`LIVE_VERIFIED` means called successfully with this atlas's keys on 2026-09-18; xAI probes ran on a Tier 0 team, Gemini probes on a free-tier key — hence `ACCOUNT_RESTRICTED` on paid-only Gemini features and on xAI's Management/Skills/Embeddings). Machine-readable twin: `generated/compatibility/cross-provider-feature-matrix.json` (same rows, `openai`/`anthropic`/`xai`/`gemini` objects per record, `providers_supporting[]`, `provider_count`).
**Sources:** the `ref` column of the JSON twin names, per row, the generated file or docs page each cell was taken from. Canonical vendor pages: https://developers.openai.com/api/docs · https://platform.claude.com/docs/en · https://docs.x.ai/developers · https://ai.google.dev/gemini-api/docs.
**Last verified:** 2026-09-18
Legend — **Portable** = the same task can be expressed with an equivalent parameter on **every provider that offers it** (mapping in the per-topic pages; a feature offered by one provider only is never portable); “— not offered” = no documented surface. Statuses use the atlas vocabulary (`DOCUMENTED`, `LIVE_VERIFIED`, `LIVE_DISCOVERED`, `BETA`, `PREVIEW`, `GA`, `LEGACY`, `DEPRECATED`, `RETIRED`, `ACCOUNT_RESTRICTED`, `UNVERIFIED`, `FAILED_VERIFICATION`, `DOCUMENTATION_INCOMPLETE`).
## Contents
1. [Core generation](#core-generation) (12 rows)
2. [Structured output & sampling](#structured-output-sampling) (10 rows)
3. [Tools — client side](#tools-client-side) (13 rows)
4. [Tools — server side / hosted](#tools-server-side-hosted) (13 rows)
5. [Multimodal input & media APIs](#multimodal-input-media-apis) (13 rows)
6. [Files & batch](#files-batch) (3 rows)
7. [Prompt caching](#prompt-caching) (2 rows)
8. [Reasoning / thinking](#reasoning-thinking) (6 rows)
9. [Context management](#context-management) (6 rows)
10. [Service tiers, limits, safety](#service-tiers-limits-safety) (8 rows)
11. [Auth, versioning, SDKs, platform](#auth-versioning-sdks-platform) (14 rows)
12. [Managed agents platforms](#managed-agents-platforms) (15 rows)
13. [Model customisation & evaluation](#model-customisation-evaluation) (5 rows)
14. [Legacy / retired surfaces](#legacy-retired-surfaces) (5 rows)
**Totals:** 125 features — on all four providers: 45 · on three: 34 · on two: 16 · on one: 26 · on none (legacy/absent everywhere): 4 · marked portable: 86.
| Coverage | Count | Features |
|---|---|---|
| All four | 45 | Primary text/multimodal generation endpoint; System / developer prompt; Message roles; Multi-turn conversation; Token counting; Model listing / catalogue; Structured outputs (JSON Schema constrained); Strict tool arguments; Sampling parameters (temperature / top_p / top_k); Stop sequences; Refusal / safety signalling in-band; Function / custom tools (you execute); Tool choice; Parallel tool calls; Web search; Code execution (sandboxed Python); Remote MCP servers; Citations / grounding metadata; Image input; PDF / document input; Files API; Batch processing (async, discounted); Prompt / context caching; Reasoning control; Reasoning visibility; Reasoning replay across turns; Context window; Max output tokens; Service tiers / processing modes; End-user identifier for abuse detection; Rate-limit tiers; Overload / capacity error; Error envelope; Authentication; Official SDKs; Webhooks (platform events); Usage & cost reporting; Cloud availability; Zero data retention / data-use terms; Managed agent harness; Create an agent session / run; Send input / events to a session; Session event stream; Client-side agent framework / CLI; Retired / deprecated beta headers, parameters and SDKs |
| anthropic + xai + gemini | 2 | Task-wide token budget (advisory); Session budgets |
| openai + anthropic + gemini | 6 | Computer use; Private-network MCP (tunnels) / agent credentials; Server-side compaction (in-flight); Hosted sandbox; Credential vaults / secrets; Artifacts / deliverables |
| openai + anthropic + xai | 12 | Deferred tool loading / tool search; Shell execution; Skills (packaged instructions + files); Stand-alone compaction request; Rate-limit headers; Request correlation; Administration API; Audit / compliance data access; Spend limits; Data residency; Self-hosted execution; Multi-agent / subagents |
| openai + xai + gemini | 14 | Legacy / secondary chat endpoint; Response storage, retrieval, deletion; Background / deferred (async) execution; JSON mode (valid JSON, no schema); File search / managed RAG; Image generation as a tool / modality inside a text call; Text-to-speech; Speech-to-text / transcription; Realtime speech-to-speech (WebSocket / WebRTC / SIP); Image generation API; Video generation API; Multipart / resumable upload; Pro / extended-compute mode; OpenAI-compatibility layer |
| anthropic + gemini | 3 | Web fetch / URL context (retrieve specific URLs); API version header / path version; Scheduled runs |
| anthropic + xai | 1 | Anthropic-compatible Messages endpoint |
| openai + anthropic | 6 | Programmatic tool calling (model writes code that calls tools); File editing tool; Cache diagnostics; Change effort mid-conversation without breaking the cache; Mid-conversation tool add/remove (cache-preserving); Beta opt-in header |
| openai + gemini | 5 | Configurable safety thresholds; Async tools (model continues while a tool runs); Container management API; Audio / video input (understanding); Embeddings |
| xai + gemini | 1 | Social / vertical search (X posts, Google Maps) |
| OpenAI only | 15 | Legacy completion (prompt-in / text-out) endpoint; Output verbosity control; Log probabilities; Free-form / grammar-constrained custom tools; Tool namespaces; Built-in SaaS connectors; Full-duplex voice with backend delegation (Live) / voice agents; Moderation endpoint; Content provenance; Embeddable chat UI & workspace agents; Fine-tuning / tuning; Evals; Graders (standalone); Stored completions / distillation; Reusable prompt templates |
| Anthropic only | 8 | Assistant prefill (continue a partial assistant turn); Server-side model fallback on refusal; Memory tool (client-side persistent memory); Browser use toolset; Advisor (consult a stronger model mid-response); Server-side context editing (clear old tool results / thinking); Outcome grading (define outcome + rubric); Server-side memory stores |
| xAI only | 2 | Per-request dollar cost in the response; Retired model ids keep resolving (redirect aliases) |
| Gemini only | 1 | Music generation |
| None | 4 | Idempotency key; Retired agent / thread APIs; Retired media models / endpoints; Retired text models |
## Core generation
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Primary text/multimodal generation endpoint**
*4/4* | Responses API: `input` (string or Item[]), `output[]` items
**Endpoint:** `POST /v1/responses`
**Params:** `model`, `input`, `instructions`, `tools`, `text`, `reasoning`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*stored by default (`store:true`, 30 days)* | Messages API: `messages[]` of content blocks, `content[]` blocks out
**Endpoint:** `POST /v1/messages`
**Params:** `model`, `messages`, `system`, `max_tokens`, `tools`, `output_config`, `thinking`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*stateless; `max_tokens` required (400 if missing)* | Responses API clone: same `input` items / `output[]` items as OpenAI (`message`, `reasoning`, `function_call`, `web_search_call`, `code_interpreter_call`, `mcp_call`…); xAI extras `max_turns`, `top_k`, `min_p`, `reasoning_effort`; `usage.cost_in_usd_ticks`
**Endpoint:** `POST /v1/responses`
**Params:** `model`, `input`, `instructions`, `tools`, `text`, `reasoning`, `max_turns`, `store`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*stored by default (30 days); `background` → 400; `metadata` → 400; every Grok call bills reasoning tokens* | `generateContent`: `contents[] {role: user\|model, parts[]}` → `candidates[].content.parts[]`; `generationConfig` holds every knob; Google calls it 'legacy' since June 2026 in favour of the **Interactions API** (`POST /v1beta/interactions`, snake_case, `model` or `agent`, `input`, `steps[]` out)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents`, `systemInstruction`, `generationConfig`, `tools`, `toolConfig`, `safetySettings`, `cachedContent`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`/v1` stable twin exists for generateContent; Interactions `/v1beta` BETA + LIVE_VERIFIED, `/v1` GA but UNVERIFIED* | yes | Four envelopes, two shapes: OpenAI and xAI share the **items** model (xAI reimplements the Responses API); Anthropic uses **content blocks**; Gemini uses **parts** inside `contents[]` with `generationConfig`. Required fields differ: Anthropic needs `max_tokens`; Gemini needs the last turn to be `user`; xAI rejects `background`/`metadata`.
*Ref:* docs/openai/responses.md · docs/anthropic/messages-api.md · docs/xai/responses.md · docs/gemini/generate-content.md · docs/gemini/interactions-api.md |
| **Legacy / secondary chat endpoint**
*3/4* | Chat Completions (`messages[]`, `choices[]`), still GA and maintained
**Endpoint:** `POST /v1/chat/completions`
**Params:** `messages`, `response_format`, `reasoning_effort`, `web_search_options`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no hosted tools except search models; list/retrieve/update/delete of stored completions* | — not offered
*none — Messages is the only chat surface* | Chat Completions, OpenAI-compatible; xAI calls it the **legacy predecessor** of `/v1/responses` (new features land on Responses first); only `function` tools; `deferred:true` + `GET /v1/chat/deferred-completion/{id}`; `reasoning_content` field
**Endpoint:** `POST /v1/chat/completions`
**Params:** `messages`, `reasoning_effort`, `response_format`, `deferred`, `prompt_cache_key`, `max_completion_tokens`
**Status:** `DOCUMENTED` · `LEGACY` · `LIVE_VERIFIED`
*`file` parts → 400 (use Responses); server tools → 422; Live Search params → 410* | OpenAI-compatibility layer `POST /v1beta/openai/chat/completions` (Bearer key required; `extra_body.google.thinking_config`, `cached_content`; unknown params silently ignored); native `generateContent` itself is now labelled legacy vs Interactions
**Endpoint:** `POST /v1beta/openai/chat/completions`
**Params:** `messages`, `reasoning_effort`, `response_format`, `extra_body.google`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*no Responses / Assistants / audio routes on the compat layer* | yes | OpenAI, xAI and Gemini all expose an OpenAI-shaped `chat/completions`; Anthropic has none. Only OpenAI's is a first-class, fully featured surface.
*Ref:* docs/comparisons/responses-vs-chat-completions.md · docs/xai/chat-completions.md · docs/gemini/openai-compatibility.md |
| **Anthropic-compatible Messages endpoint**
*2/4* | — not offered | the native surface
**Endpoint:** `POST /v1/messages`
**Params:** `messages`, `system`, `max_tokens`, `tools`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `POST /v1/messages` accepting Anthropic shapes (`system`, `messages`, `tools[{name,description,input_schema}]`, `tool_choice {auto\|any\|tool}`, `thinking` blocks with empty `signature`) — **fully deprecated** by xAI ('migrate to Responses or gRPC'); Bearer auth, `anthropic-version` ignored; `/v1/complete` twin RETIRED (400)
**Endpoint:** `POST /v1/messages`
**Params:** `messages`, `system`, `max_tokens`, `tools`, `tool_choice`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED`
*`top_k` 400, `stop_sequences` 400 on reasoning models, `document` blocks 422, no count_tokens (404), no batches, `cache_control` ignored* | — not offered | yes | xAI is the only third party exposing the Anthropic Messages wire format, and is retiring it.
*Ref:* docs/xai/messages-compat.md · parameters.json (xai POST /v1/messages) |
| **Legacy completion (prompt-in / text-out) endpoint**
*1/4* | `gpt-3.5-turbo-instruct`, `davinci-002`, `babbage-002` only; shutdowns 2026-09-28
**Endpoint:** `POST /v1/completions`
**Params:** `prompt`, `suffix`, `max_tokens`
**Status:** `DOCUMENTED` · `LEGACY` · `LIVE_VERIFIED` | documented as Legacy but every live call returns 400 'has been deprecated'
**Endpoint:** `POST /v1/complete`
**Params:** `prompt`, `max_tokens_to_sample`
**Status:** `DOCUMENTED` · `LEGACY` · `DEPRECATED` · `FAILED_VERIFICATION`
*removed from Python SDK v1* | `POST /v1/completions` (OpenAI shape) and `POST /v1/complete` (Anthropic shape) both answer 400 'Raw sampling is not supported for reasoning models' for every live model, incl. the non-reasoning grok-4.20
**Endpoint:** `POST /v1/completions`
**Params:** `prompt`, `max_tokens`
**Status:** `DOCUMENTED` · `LEGACY` · `RETIRED`
*gRPC `Sample` service still lists raw sampling* | PaLM-era `:generateText` / `:generateMessage` remain in the v1beta discovery document but return 404 / 501; only `models/aqa:generateAnswer` still answers
**Endpoint:** `POST /v1beta/models/{model}:generateText`
**Status:** `DOCUMENTED` · `LEGACY` · `FAILED_VERIFICATION`
*`legacy-palm` family: 7 endpoints* | no | Every provider still documents a prompt-completion route; only OpenAI's answers, and it shuts down 2026-09-28.
*Ref:* docs/openai/completions-legacy.md · docs/anthropic/text-completions-legacy.md · docs/xai/legacy-completions.md · docs/gemini/legacy-palm-methods.md |
| **System / developer prompt**
*4/4* | `instructions` (per request, not carried by `previous_response_id`) or `developer`/`system` role message items
**Endpoint:** `POST /v1/responses`
**Params:** `instructions`, `input[](message).role`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`instructions` participate in the cache prefix* | top-level `system` (string or text blocks with `cache_control`/`citations`); mid-conversation `role:system` messages for tool changes/effort (beta)
**Endpoint:** `POST /v1/messages`
**Params:** `system`, `system[].cache_control`, `messages[].output_config.effort`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*schema lists a `system` role but docs say use the top-level field* | `instructions` (Responses) **or** a `system` / `developer` role message; Chat: `system` role; `/v1/messages`: `system` string or text blocks
**Endpoint:** `POST /v1/responses`
**Params:** `instructions`, `input[](message).role`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`instructions` + `previous_response_id` → 400 (the previous system prompt is reused)* | `systemInstruction` (Content, text parts only, role ignored) — not a turn; Interactions: `system_instruction` string (must be resent when chaining)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `systemInstruction`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*counted in `promptTokenCount`; cacheable inside `cachedContents`* | yes | All four keep the system prompt outside the turn list. OpenAI/xAI accept it either as a field or as a role item; Anthropic and Gemini only as a field.
*Ref:* parameters.json (instructions / system / systemInstruction) |
| **Message roles**
*4/4* | `user`, `assistant`, `system`, `developer` (+ `phase: commentary\|final_answer` on assistant)
**Endpoint:** `POST /v1/responses`
**Params:** `input[](message).role`, `input[](message).phase`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `user`, `assistant`; consecutive same-role turns are merged; ≤100,000 messages
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].role`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`system` role reserved for beta mid-conversation blocks* | `user`, `assistant`, `system`, `developer` (Responses); Chat adds `tool` (`tool_call_id`) and legacy `function`; no ordering constraint
**Endpoint:** `POST /v1/responses`
**Params:** `input[](message).role`, `messages[].role`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`developer` accepted live on Chat although not in the spec* | `user`, `model` only (omitted role = user); **last turn must be `user`** (400 'Requests ending with a model turn are not supported'); function results go in a `user` Content of `functionResponse` parts
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].role`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Interactions steps use `role: user\|model` too* | yes | The assistant role is spelled `model` on Gemini; Anthropic enforces alternation by merging; Gemini enforces a trailing user turn; OpenAI/xAI accept any order.
*Ref:* docs/anthropic/messages-api.md §1 · docs/gemini/generate-content.md |
| **Assistant prefill (continue a partial assistant turn)**
*1/4* | no prefill semantics; an assistant item is history only
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | end `messages` with an `assistant` turn → model continues it. **Deprecated**: 400 on Claude 4.6+ / Fable / Mythos, incompatible with thinking and structured outputs
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].content[] (assistant prefill)`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED`
*live: Haiku 4.5 OK, Sonnet 5 → 400* | assistant history items / `assistant` role accepted, but no documented continuation semantics
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | impossible: a request ending with a `model` turn is rejected (400)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no | Only Anthropic ever offered prefill, and only its 4.5 models still honour it; Gemini rejects the pattern outright.
*Ref:* docs/anthropic/messages-api.md · deprecations.json · docs/gemini/generate-content.md |
| **Multi-turn conversation**
*4/4* | three modes: manual replay of `output[]` items, `previous_response_id` (server keeps chain), `conversation` (Conversations API)
**Endpoint:** `POST /v1/responses`
**Params:** `previous_response_id`, `conversation`, `store`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`previous_response_id` and `conversation` are mutually exclusive* | manual replay only: resend full `messages[]` each call (incl. `thinking`/`tool_use`/`tool_result` blocks)
**Endpoint:** `POST /v1/messages`
**Params:** `messages`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*server-side state exists only in Managed Agents sessions* | manual replay **or** `previous_response_id` (server rehydrates the whole agentic trajectory incl. reasoning and tool outputs; follow-ups may change tools/model); Chat: replay incl. `reasoning_content`
**Endpoint:** `POST /v1/responses`
**Params:** `previous_response_id`, `store`, `input`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no Conversations API; `x-grok-conv-id` / `prompt_cache_key` are cache-routing keys, not stored conversations* | `generateContent`: manual replay of `contents[]` (echo `thoughtSignature` on Gemini 3 function calls); **Interactions**: `previous_interaction_id` chains stored interactions (`store:true` default; only history is carried — resend tools/system/config)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents`, `previous_interaction_id`, `store`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*chaining on an `in_progress` interaction → 400* | yes | Portable pattern = manual replay everywhere; OpenAI, xAI and Gemini (Interactions) add server-side chaining by id.
*Ref:* docs/comparisons/state-management.md |
| **Response storage, retrieval, deletion**
*3/4* | `store:true` default → `GET/DELETE /v1/responses/{id}`, `GET …/input_items`; 30-day retention
**Endpoint:** `GET /v1/responses/{response_id}`
**Params:** `store`, `include`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`store:false` → 404 on retrieve, reasoning returned as `encrypted_content`* | no response store; nothing to retrieve after the HTTP response
**Status:** `DOCUMENTED`
*Batches results are retrievable 29 days* | `store:true` default → `GET/DELETE /v1/responses/{id}`, `GET …/input_items` (limit 1–100, order, after); 30-day retention; ZDR teams cannot store
**Endpoint:** `GET /v1/responses/{response_id}`
**Params:** `store`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` · `LIVE_DISCOVERED`
*LIVE_DISCOVERED: GET still 200 for a `store:false` id* | Interactions: `store:true` default → `GET /v1beta/interactions/{id}` (full timeline incl. `user_input`), `DELETE`, `POST …/cancel`; retention 55 days paid (AI Studio 7/14/28/55), 1 day free; `store:false` → no `id`, stateless. `generateContent` stores nothing (per-request `store` logging flag LIVE_DISCOVERED)
**Endpoint:** `GET /v1beta/interactions/{id}`
**Params:** `store`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*`GET /v1beta/interactions` (list) → 404* | yes | Three stateful-by-default surfaces (OpenAI Responses, xAI Responses, Gemini Interactions) vs Anthropic's stateless Messages.
*Ref:* docs/openai/responses.md §1 · docs/xai/responses.md · docs/gemini/interactions-api.md |
| **Background / deferred (async) execution**
*3/4* | `background:true` → `status:queued`, poll `GET`, `POST …/cancel`, resumable stream `?stream=true&starting_after=N`
**Endpoint:** `POST /v1/responses`
**Params:** `background`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*not available over WebSocket; not in EU region* | no per-request background mode; use Message Batches (≤24 h) or Managed Agents sessions
**Endpoint:** `POST /v1/messages/batches`
**Status:** `DOCUMENTED` | Chat Completions only: `deferred:true` → `{request_id}`; `GET /v1/chat/deferred-completion/{request_id}` → 202 while pending, 200 when done (docs: retrievable once within 24 h; live: second GET also 200)
**Endpoint:** `GET /v1/chat/deferred-completion/{request_id}`
**Params:** `deferred`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Responses `background` → 400 'Argument not supported'* | Interactions `background:true` (requires `store`) → `queued`/`in_progress`; poll `GET /v1beta/interactions/{id}` or resume the stream with `?stream=true&last_event_id=`; `POST …/cancel`; mandatory for Deep Research (≤60 min); `webhook_config` for completion callbacks
**Endpoint:** `POST /v1beta/interactions`
**Params:** `background`, `stream`, `webhook_config`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*Veo/Batch use long-running `Operation`s instead* | yes | OpenAI, xAI (Chat only) and Gemini (Interactions) offer per-request async; Anthropic only batches.
*Ref:* docs/openai/responses.md §5 · docs/xai/deferred-completions.md · docs/gemini/interactions-api.md |
| **Token counting**
*4/4* | count input tokens of a Responses payload (free)
**Endpoint:** `POST /v1/responses/input_tokens`
**Params:** `model`, `input`, `tools`, `instructions`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*supports images/files* | count tokens of a Messages payload (free; separate RPM bucket 5k/10k/20k)
**Endpoint:** `POST /v1/messages/count_tokens`
**Params:** `model`, `messages`, `system`, `tools`, `thinking`, `output_config.format`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*server tools other than advisor → 400; `mcp_servers` rejected* | `POST /v1/tokenize-text {model, text}` → `token_ids[{token_id, string_token, token_bytes}]` (tokenizer, not a request counter); free
**Endpoint:** `POST /v1/tokenize-text`
**Params:** `model`, `text`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no per-request counter; `/v1/messages/count_tokens` → 404* | `:countTokens` with `{contents}` or `{generateContentRequest:{model, contents, systemInstruction, tools, cachedContent, generationConfig}}` → `totalTokens`, `promptTokensDetails[]`, `cachedContentTokenCount`; free
**Endpoint:** `POST /v1beta/models/{model}:countTokens`
**Params:** `contents`, `generateContentRequest`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*SDKs refuse `system_instruction`/`tools` here although REST accepts them* | yes | Free everywhere; xAI tokenises raw text only (no tools/images), the other three count a full request.
*Ref:* docs/anthropic/token-counting.md · docs/openai/multimodal-input.md · docs/xai/index.md · docs/gemini/token-counting.md |
| **Model listing / catalogue**
*4/4* | list/retrieve (`shutdown_date` field on deprecated ids); some aliases (e.g. `gpt-5.6`) 404 on GET but work on POST
**Endpoint:** `GET /v1/models`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*136 ids live* | list/retrieve with `capabilities` block (thinking, effort, structured_outputs, context_management…)
**Endpoint:** `GET /v1/models`
**Params:** `limit`, `before_id`, `after_id`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*11 ids live; invite-only Mythos ids → 404* | `GET /v1/models` (OpenAI shape, 12 ids) plus typed catalogues `GET /v1/language-models`, `/v1/image-generation-models`, `/v1/video-generation-models`, `/v1/embedding-models` with **live price ticks**, aliases, `input_modalities`, fingerprint; retired slugs redirect (`grok-3` → grok-4.3 object)
**Endpoint:** `GET /v1/language-models`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*voice models absent from the catalogue; `/v1/embedding-models` → `{models: []}` for this team* | `GET /v1beta/models` (58 ids: `inputTokenLimit`, `outputTokenLimit`, `supportedGenerationMethods`, `thinking`, `temperature`/`topP`/`topK` defaults, `version`) and `GET /v1beta/models/{model}`; `/v1/models` lists only 22 stable ids
**Endpoint:** `GET /v1beta/models`
**Params:** `pageSize`, `pageToken`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*agents (deep-research, antigravity) appear as models; shut-down previews still listed* | yes | xAI publishes prices in the catalogue, Anthropic publishes capability flags, Gemini publishes token limits and generation methods, OpenAI publishes shutdown dates.
*Ref:* sources/*/models-api-raw.json |
## Structured output & sampling
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Structured outputs (JSON Schema constrained)**
*4/4* | `text.format {type:json_schema, name, schema, strict, description}` (flat)
**Endpoint:** `POST /v1/responses`
**Params:** `text.format`, `text.format(json_schema).strict`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Chat: `response_format.json_schema{…}` wrapper; refusal → `refusal` content part* | `output_config.format {type:json_schema, schema}` (GA, no beta header; legacy `output_format` → 400)
**Endpoint:** `POST /v1/messages`
**Params:** `output_config.format`, `output_config.format.schema`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*grammar compiled once (24 h cache), first call slower; `stop_reason: refusal` possible* | `text.format {type:json_schema, name, schema, strict, description}` (Responses) / `response_format {type:json_schema, json_schema:{name, strict, schema}}` (Chat); Draft 2020-12 preferred; `additionalProperties` defaults false; formats date/time/email/uuid/uri enforced; `pattern` ECMA subset
**Endpoint:** `POST /v1/responses`
**Params:** `text.format`, `response_format`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`strict` accepted and ignored (always strict); all Grok 4 models* | `generationConfig.responseMimeType: application/json` + `responseJsonSchema` (JSON Schema) or legacy `responseSchema` (OpenAPI subset, `propertyOrdering`); new canonical `responseFormat.text {mimeType: APPLICATION_JSON, schema}`; enum mode `text/x.enum`; XML/YAML mime types accepted; Interactions `response_format {type:text, mime_type, schema}`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.responseMimeType`, `generationConfig.responseJsonSchema`, `generationConfig.responseSchema`, `generationConfig.responseFormat`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*values are syntactically valid but not semantically validated; `maxOutputTokens` can truncate the JSON; SO + tools = Gemini 3 preview* | yes | Same task on all four. Schema dialects differ: OpenAI/xAI require `additionalProperties:false` semantics (xAI defaults it), Anthropic rejects numeric/string constraints, Gemini ignores unsupported keywords silently and offers an OpenAPI-style alternative.
*Ref:* docs/openai/structured-outputs.md · docs/anthropic/structured-outputs.md · docs/xai/structured-outputs.md · docs/gemini/structured-outputs.md |
| **Strict tool arguments**
*4/4* | `tools[type=function].strict:true` (Responses omits → tries strict then falls back)
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=function].strict`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `tools[].strict:true` grammar-constrained `tool_use.input`; ≤20 strict tools/request
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].strict`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*not on toolsets / mcp_toolset / programmatic callers* | tool `parameters` are **always** strictly enforced ('strict flag implicitly true'); explicit `strict` accepted and ignored
**Endpoint:** `POST /v1/responses`
**Params:** `tools[].strict`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `toolConfig.functionCallingConfig.mode: VALIDATED` (schema-validated constrained decoding; default when built-ins or structured output are combined) or `ANY` (forced, constrained)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `toolConfig.functionCallingConfig.mode`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no per-tool flag; `ANY` may reject very large/deep schemas* | yes | Per-tool flag on OpenAI/Anthropic, always-on on xAI, a request-level mode on Gemini.
*Ref:* tools.json · docs/tools/gemini/function-calling.md |
| **JSON mode (valid JSON, no schema)**
*3/4* | `text.format {type:json_object}`; prompt must mention JSON (400 otherwise)
**Endpoint:** `POST /v1/responses`
**Params:** `text.format(json_object)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*legacy* | — not offered | `text.format {type:json_object}` / `response_format {type:json_object}`
**Endpoint:** `POST /v1/responses`
**Params:** `text.format(json_object)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `responseMimeType: application/json` without a schema
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.responseMimeType`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | Anthropic only offers schema-constrained output.
*Ref:* docs/openai/structured-outputs.md · docs/gemini/structured-outputs.md |
| **Output verbosity control**
*1/4* | `text.verbosity: low\|medium\|high`
**Endpoint:** `POST /v1/responses`
**Params:** `text.verbosity`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no direct knob; `output_config.effort` shapes length indirectly
**Status:** `DOCUMENTED` | no verbosity parameter (`text.format` only)
**Status:** `DOCUMENTED` | no verbosity parameter; `thinkingLevel` and `maxOutputTokens` only
**Status:** `DOCUMENTED` | no | OpenAI-only knob.
*Ref:* parameters.json |
| **Sampling parameters (temperature / top_p / top_k)**
*4/4* | `temperature` 0–2, `top_p`; rejected by reasoning models
**Endpoint:** `POST /v1/responses`
**Params:** `temperature`, `top_p`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no `top_k`* | `temperature` 0–1, `top_p`, `top_k` — **DEPRECATED**: 400 when non-default on Claude 4.7+ / Fable / Mythos; Python SDK v1 removed the kwargs
**Endpoint:** `POST /v1/messages`
**Params:** `temperature`, `top_p`, `top_k`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED` | `temperature` 0–2, `top_p`, plus xAI-specific `top_k` (≥1) and `min_p` (0–1); accepted on reasoning models (echo 0.7/0.95); Chat `seed` → `system_fingerprint`
**Endpoint:** `POST /v1/responses`
**Params:** `temperature`, `top_p`, `top_k`, `min_p`, `seed`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`presence_penalty`/`frequency_penalty`/`stop` → 400 on reasoning models* | `generationConfig.temperature` 0–2 (default 1.0), `topP` (0.95), `topK` (64), `seed`; **DEPRECATED** guidance since 2026-07-21: keep defaults on Gemini 3.x (temperature < 1 can cause looping); `presencePenalty`/`frequencyPenalty` → 400 'not enabled'
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.temperature`, `generationConfig.topP`, `generationConfig.topK`, `generationConfig.seed`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED`
*`candidateCount` > 1 → 400 on 3.x* | yes | Sampling knobs are shrinking on all frontier lines: rejected by OpenAI reasoning models and Claude 4.7+, deprecated-by-guidance on Gemini 3.x, still fully accepted on Grok.
*Ref:* deprecations.json (anthropic, gemini api_features) · parameters.json |
| **Stop sequences**
*4/4* | Chat Completions `stop` (≤4); **not** a Responses parameter
**Endpoint:** `POST /v1/chat/completions`
**Params:** `stop`
**Status:** `DOCUMENTED` | `stop_sequences[]` → `stop_reason: stop_sequence` + `stop_sequence`
**Endpoint:** `POST /v1/messages`
**Params:** `stop_sequences`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | Chat `stop` (≤4) and `/v1/messages` `stop_sequences` — **400 on reasoning models**; not on Responses
**Endpoint:** `POST /v1/chat/completions`
**Params:** `stop`, `stop_sequences`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*usable only on grok-4.20-0309-non-reasoning* | `generationConfig.stopSequences[]` (≤5; 6 → 400); Interactions `generation_config.stop_sequences`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.stopSequences`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | Absent from both Responses APIs; Gemini allows 5, the others 4.
*Ref:* parameters.json |
| **Log probabilities**
*1/4* | `top_logprobs` + `include: ["message.output_text.logprobs"]`
**Endpoint:** `POST /v1/responses`
**Params:** `top_logprobs`, `include`
**Status:** `DOCUMENTED` | — not offered | `logprobs`/`top_logprobs` (0–8) accepted but **silently ignored** on grok-4.20 and newer (DEPRECATED); `include: message.output_text.logprobs` ignored
**Endpoint:** `POST /v1/chat/completions`
**Params:** `logprobs`, `top_logprobs`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED` | `responseLogprobs` / `logprobs` (0–20) in the schema but `400 Logprobs is not enabled for this model` on every current model
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.responseLogprobs`, `generationConfig.logprobs`
**Status:** `DOCUMENTED` · `FAILED_VERIFICATION` | no | Only OpenAI still returns logprobs; xAI and Gemini keep the parameters as dead compatibility fields.
*Ref:* parameters.json |
| **Refusal / safety signalling in-band**
*4/4* | `output[].content[] {type: refusal}` (structured outputs) / `status: incomplete` + `incomplete_details.reason: content_filter`; HTTP 403 `misalignment_policy_violation`, `cyber_policy`
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | HTTP 200 with `stop_reason: refusal` + `stop_details {category, explanation}`; beta `fallbacks` re-runs on another model
**Endpoint:** `POST /v1/messages`
**Params:** `fallbacks`, `fallback_credit_token`
**Status:** `DOCUMENTED` · `BETA`
*categories: cyber, bio, frontier_llm, reasoning_extraction, general_harms* | Chat `message.refusal` field; `respect_moderation` flag on image/video results; usage-guideline violations are **billed** (+ $0.05 fee when caught pre-generation on Responses); no error code catalogue for refusals
**Endpoint:** `POST /v1/chat/completions`
**Status:** `DOCUMENTED` | HTTP 200 with **no candidates** + `promptFeedback.blockReason` (SAFETY\|OTHER\|BLOCKLIST\|PROHIBITED_CONTENT\|IMAGE_SAFETY) for prompt blocks; `candidates[].finishReason` SAFETY\|RECITATION\|SPII\|IMAGE_SAFETY\|… (21 values) + `safetyRatings[]` for output blocks; thresholds via `safetySettings[]`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `safetySettings`, `candidates[].finishReason`, `promptFeedback.blockReason`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Interactions mirror finishReason as snake_case error codes* | yes | All four signal in-band with different shapes; only Gemini lets the caller tune thresholds; only xAI charges for violating requests.
*Ref:* docs/anthropic/stop-reasons.md · docs/openai/safety.md · docs/gemini/safety.md · docs/xai/pricing.md §5 |
| **Server-side model fallback on refusal**
*1/4* | — not offered | `fallbacks` parameter (+ `fallback_credit_token` billing credit)
**Endpoint:** `POST /v1/messages`
**Params:** `fallbacks`, `fallback_credit_token`
**Status:** `DOCUMENTED` · `BETA`
*beta `server-side-fallback-2026-07-01`, `fallback-credit-2026-07-01`* | — not offered | — not offered | no | Anthropic-only.
*Ref:* generated/fragments/headers/anthropic-beta-headers.json |
| **Configurable safety thresholds**
*2/4* | inline `moderation {model, policy}` on Responses/Chat (policy selection, not thresholds)
**Endpoint:** `POST /v1/responses`
**Params:** `moderation`
**Status:** `DOCUMENTED` | — not offered | — not offered | `safetySettings[] {category: HARM_CATEGORY_HARASSMENT\|HATE_SPEECH\|SEXUALLY_EXPLICIT\|DANGEROUS_CONTENT\|JAILBREAK (CIVIC_INTEGRITY deprecated → `enableEnhancedCivicAnswers`), threshold: OFF\|BLOCK_NONE\|BLOCK_ONLY_HIGH\|BLOCK_MEDIUM_AND_ABOVE\|BLOCK_LOW_AND_ABOVE}`; default Off on 2.5/3.x; `safetyRatings[]` returned when a threshold is set
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `safetySettings`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Interactions: custom safety settings not supported* | no | Only Gemini exposes per-category blocking thresholds; OpenAI selects a moderation policy.
*Ref:* docs/gemini/safety.md · docs/openai/moderation.md |
## Tools — client side
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Function / custom tools (you execute)**
*4/4* | `tools[type=function] {name, description, parameters, strict, output_schema, async, defer_loading, allowed_callers}` → `function_call` item → you reply `function_call_output {call_id, output}`
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=function]`, `input[](function_call_output)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*38 compatible models listed* | `tools[] {name, description, input_schema, strict, input_examples, cache_control, defer_loading, allowed_callers}` → `tool_use` block → you reply user message of `tool_result {tool_use_id, content, is_error}`
**Endpoint:** `POST /v1/messages`
**Params:** `tools`, `messages[].content[] (tool_result)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*13 compatible models* | Responses `tools[type=function] {name, description, parameters}` → `function_call {call_id: call-…, name, arguments}` → `function_call_output`; Chat `tools[{type:function, function:{…}}]` → `message.tool_calls[]` → `{role: tool, tool_call_id}`; ≤350 tools; parameters always strict; a client-side call ends the agentic request (fresh `max_turns` budget on the follow-up)
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=function]`, `input[](function_call_output)`, `parallel_tool_calls`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*all 7 Grok text models; also inside the voice session and `/v1/messages`* | `tools[].functionDeclarations[] {name, description, parameters \| parametersJsonSchema, response \| responseJsonSchema, behavior}` → `parts[].functionCall {name, args, id}` (+ mandatory `thoughtSignature` on Gemini 3) → you reply a `user` Content of `functionResponse {name, id, response, parts[] (multimodal)}`; max 512 declarations
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].functionDeclarations`, `contents[].parts[].functionResponse`, `toolConfig.functionCallingConfig`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Interactions: `{type: function, name, description, parameters}` + `function_call`/`function_result` steps; Live: `toolCall`/`toolResponse` messages* | yes | Same loop on all four. Arguments are a JSON **string** on OpenAI/xAI and an **object** on Anthropic (`input`) and Gemini (`args`); results are a top-level item (OpenAI/xAI), a leading `tool_result` block in a user turn (Anthropic) or `functionResponse` parts in a user Content (Gemini). Gemini 3 additionally requires echoing the `thoughtSignature` of the first call of each step (400 / `MISSING_THOUGHT_SIGNATURE`).
*Ref:* docs/comparisons/tool-execution.md · docs/tools/xai/function-calling.md · docs/tools/gemini/function-calling.md |
| **Free-form / grammar-constrained custom tools**
*1/4* | `tools[type=custom] {format: {type:text} \| {type:grammar, syntax: lark\|regex, definition}}`
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=custom].format`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | — not offered | no `custom` tool type in the deserializer (422 'unknown variant'); no grammar mode
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | — not offered | no | OpenAI-only.
*Ref:* tools.json · docs/xai/structured-outputs.md |
| **Tool choice**
*4/4* | `tool_choice`: `auto\|none\|required` string, `{type:function,name}`, hosted `{type:web_search\|mcp\|shell\|apply_patch\|…}`, `{type:allowed_tools, mode, tools[]}`
**Endpoint:** `POST /v1/responses`
**Params:** `tool_choice`, `tool_choice(allowed_tools)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `tool_choice {type: auto\|any\|tool\|none, name?, disable_parallel_tool_use?}`
**Endpoint:** `POST /v1/messages`
**Params:** `tool_choice.type`, `tool_choice.disable_parallel_tool_use`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`any`/`tool` → 400 on Fable 5.1 / Mythos 5.1 and with manual thinking* | `tool_choice`: `auto\|none\|required` or `{type:function, name}` (Responses) / `{type:function, function:{name}}` (Chat); `/v1/messages`: `{type: auto\|any\|tool}` (`disable_parallel_tool_use` → 400); no `allowed_tools`, no forcing of server tools
**Endpoint:** `POST /v1/responses`
**Params:** `tool_choice`, `tool_choice.type`, `tool_choice.name`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `toolConfig.functionCallingConfig {mode: AUTO\|ANY\|NONE\|VALIDATED, allowedFunctionNames[]}` (ANY + names = forced subset); Interactions `generation_config.tool_choice: auto\|any\|none\|validated` or `{allowed_tools:{mode, tools[]}}`; built-in tools cannot be forced
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `toolConfig.functionCallingConfig.mode`, `toolConfig.functionCallingConfig.allowedFunctionNames`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Live `setup.toolConfig` → close 1007* | yes | `required` ≈ `any` ≈ `ANY`; only OpenAI and Gemini (`allowedFunctionNames` / Interactions `allowed_tools`) can restrict to a subset; only OpenAI can force a hosted tool.
*Ref:* parameters.json |
| **Parallel tool calls**
*4/4* | `parallel_tool_calls` (default true); hosted tools never batched with functions
**Endpoint:** `POST /v1/responses`
**Params:** `parallel_tool_calls`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | default on (Claude 4+); `tool_choice.disable_parallel_tool_use:true` to force ≤1; all results in one user message
**Endpoint:** `POST /v1/messages`
**Params:** `tool_choice.disable_parallel_tool_use`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `parallel_tool_calls` (default true; `false` = at most one call) on Chat and Responses
**Endpoint:** `POST /v1/responses`
**Params:** `parallel_tool_calls`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`/v1/messages` `disable_parallel_tool_use` → 400* | several `functionCall` parts in one `model` Content; answer **all** of them in one `user` Content (interleaving FC1,FR1,FC2 → 400); only the first call carries the thought signature; no on/off switch
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].parts[].functionCall`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`compositional_function_calling` chains calls across turns* | yes | Default-on everywhere; Gemini has no switch to disable it.
*Ref:* docs/openai/tool-loop.md · docs/tools/anthropic/tool-use-loop.md · docs/tools/gemini/function-calling.md |
| **Tool namespaces**
*1/4* | `tools[type=namespace] {name, description, tools[]}`; calls carry `namespace`
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=namespace]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no namespaces; toolsets (`computer_toolset_20260801`, `mcp_toolset`) group Anthropic-defined members only
**Status:** `DOCUMENTED` | — not offered | — not offered | no | OpenAI-only.
*Ref:* tools.json |
| **Deferred tool loading / tool search**
*3/4* | `tools[type=tool_search] {execution: server\|client}` + `defer_loading:true` on function/custom/mcp; `tool_search_call`/`tool_search_output` items; `additional_tools` input item
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=tool_search]`, `tools[type=function].defer_loading`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*gpt-5.4+ only; gpt-5.4-nano lacks it* | `tool_search_tool_regex_20251119` / `tool_search_tool_bm25_20251119` (server) + `defer_loading:true` (≤10,000 tools); `tool_reference` blocks expanded server-side; client-side search via `tool_reference` in `tool_result`
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].defer_loading`, `tools[].type`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*GA no header; not cacheable together with `cache_control`* | `tools[type=tool_search] {execution}` + `defer_loading:true` on function/mcp tools → `tool_search_call {arguments:{query, limit}}` / `tool_search_output {tools[]}` — **alpha**: 403 'only available for alpha users'
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=tool_search]`, `tools[].defer_loading`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*listed as compatible with all 7 text models* | — not offered | yes | OpenAI-shaped on xAI (gated), two algorithms on Anthropic; Gemini has no deferred loading (best practice: keep 10–20 active declarations).
*Ref:* docs/tools/openai/tool-search.md · docs/tools/anthropic/tool-search.md · tools.json (xai tool_search) |
| **Programmatic tool calling (model writes code that calls tools)**
*2/4* | `tools[type=programmatic_tool_calling]` + `allowed_callers:["programmatic"]`; `program`/`program_output` items; nested calls `caller.type: program`; JavaScript in isolated V8
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=programmatic_tool_calling]`, `tools[type=function].allowed_callers`
**Status:** `DOCUMENTED` · `UNVERIFIED`
*no compatible model list in tools.json* | `code_execution_20260120+` + `allowed_callers:["code_execution_20260120"]`; Python `await tool({...})` in the sandbox; `caller {type, tool_id}`; reply requires top-level `container`
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].allowed_callers`, `container`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*not Haiku 4.5 (400)* | — not offered | not a generateContent feature; nearest: compositional function calling (chained calls across turns) and code execution over `functionResponse` data; Antigravity agents script tools inside their sandbox
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` | yes | OpenAI and Anthropic only (JS vs Python).
*Ref:* docs/tools/anthropic/programmatic-tool-calling.md · docs/openai/tool-loop.md §6 · docs/tools/gemini/function-calling.md |
| **Async tools (model continues while a tool runs)**
*2/4* | `tools[type=function].async:true` + wait tool / task handles (GPT-6 Astra+)
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=function].async`, `input[](function_call).async`
**Status:** `DOCUMENTED` | — not offered | — not offered | Live API only: `functionDeclarations[].behavior: NON_BLOCKING` (default on gemini-3.8-live) + `toolResponse.functionResponses[].scheduling: INTERRUPT\|WHEN_IDLE\|SILENT`, `willContinue`; `generateContent` → 400
**Endpoint:** `WSS BidiGenerateContent`
**Params:** `setup.tools[].functionDeclarations[].behavior`, `toolResponse.functionResponses[].scheduling`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*gemini-3.8-live-extended-thinking: async only* | yes | Different scopes: OpenAI in text Responses, Gemini in the voice Live session.
*Ref:* parameters.json · docs/gemini/live-api.md |
| **Shell execution**
*3/4* | `tools[type=shell]` hosted (`environment.type: container_auto\|container_reference`) **or** local (`type: local` — you run commands); legacy `local_shell` → 400
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=shell].environment`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*gpt-5.2+, codex, gpt-6-astra* | `bash_20250124` client tool (you run a persistent bash) **or** server `bash_code_execution` sub-tool of `code_execution_20250825+`
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].type`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*all current models* | `tools[type=shell] {environment:{type: local, skills:[{name, description, path}]}}` → `shell_call` → you reply `shell_call_output {stdout, stderr, outcome}`; **local only** (no hosted container variant); hosted Python via `code_interpreter`
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=shell].environment`, `input[](shell_call_output)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*accepted live, not invoked; Grok Build CLI runs its own sandbox* | no shell tool on `generateContent`; the **Antigravity** managed agent runs bash/python/node `code_execution` and file tools inside its Linux sandbox (Interactions API only, PREVIEW)
**Endpoint:** `POST /v1beta/interactions`
**Params:** `agent_config(antigravity)`
**Status:** `DOCUMENTED` · `PREVIEW`
*`gemini-3.1-pro-preview-customtools` is tuned for bash-style custom tools you define yourself* | yes | OpenAI has hosted + local, Anthropic hosted (code execution) + local (bash), xAI local only, Gemini only inside its managed agent.
*Ref:* docs/tools/openai/shell.md · docs/tools/anthropic/bash.md · docs/xai/skills-api.md · docs/gemini/interactions-api.md |
| **File editing tool**
*2/4* | `tools[type=apply_patch]` → `apply_patch_call {operation: create_file\|update_file\|delete_file, diff}`; you reply `apply_patch_call_output`
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=apply_patch]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*undocumented SSE events `response.apply_patch_call_operation_diff.*` (LIVE_DISCOVERED)* | `text_editor_20250728` (`name: str_replace_based_edit_tool`, `max_characters`) → commands view/str_replace/create/insert; server variant `text_editor_code_execution`
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].max_characters`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*older 20250429/20250124 → 400 on current models* | — not offered | no editor tool on `generateContent`; Antigravity 09-2026 built-ins `write_to_file`, `replace_file_content`, `view_file`, `list_dir`, `find_by_name`, `grep_search` (agent sandbox only)
**Endpoint:** `POST /v1beta/interactions`
**Status:** `DOCUMENTED` · `PREVIEW`
*05-2026 tool names deprecated → 2026-10-05* | yes | Diff-based (OpenAI) vs command-based (Anthropic); Gemini only inside Antigravity; none on xAI.
*Ref:* docs/tools/openai/apply-patch.md · docs/tools/anthropic/text-editor.md · docs/gemini/interactions-api.md |
| **Memory tool (client-side persistent memory)**
*1/4* | none in Responses; Agents API sessions persist items but no memory tool
**Status:** `DOCUMENTED` | `memory_20250818` client tool (you store files under `/memories`)
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].type`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Managed Agents add server-side memory stores* | — not offered | — not offered | no | Anthropic-only.
*Ref:* docs/tools/anthropic/memory.md |
| **Computer use**
*3/4* | `tools[type=computer]` (current, gpt-5.4+/gpt-6) or `computer_use_preview` (+ `computer-use-preview` model, RETIRED); `computer_call {action\|actions[], pending_safety_checks}` → `computer_call_output {computer_screenshot, acknowledged_safety_checks}`
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=computer]`, `tools[type=computer_use_preview]`
**Status:** `DOCUMENTED` · `PREVIEW` · `ACCOUNT_RESTRICTED`
*`computer` UNVERIFIED live* | `computer_toolset_20260801` (GA, no header, Fable/Mythos/Opus 5/Sonnet 5/Opus 4.8) or beta `computer_20251124` (`computer-use-2025-11-24`) / `computer_20250124`; member tools screenshot/zoom/click…; batch actions
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].display_width_px`, `tools[].enable_zoom`, `tools[].configs`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` · `BETA`
*~4,500-token toolset definition* | — not offered | `tools[].computerUse {environment: ENVIRONMENT_BROWSER\|MOBILE\|DESKTOP, excludedPredefinedFunctions[], enablePromptInjectionDetection, disabledSafetyPolicies[]}` → predefined `functionCall`s (`click`, `type`, `scroll`, `navigate`, `open_app`…; coordinates 0–999) → you reply `functionResponse` with a screenshot; `safety_decision: require_confirmation` → `safety_acknowledgement`; Interactions `{type: computer_use, environment: browser}`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].computerUse`, `tools[].computerUse.environment`
**Status:** `DOCUMENTED` · `PREVIEW` · `ACCOUNT_RESTRICTED`
*gemini-3.8-flash recommended; legacy `gemini-2.5-computer-use-preview-10-2025` browser-only; no free tier* | yes | Three client-executed screenshot loops (OpenAI, Anthropic, Gemini); none on xAI. Gemini normalises coordinates to 0–999 and adds mobile/desktop environments.
*Ref:* docs/tools/openai/computer-use.md · docs/tools/anthropic/computer-use.md · docs/tools/gemini/computer-use.md |
| **Browser use toolset**
*1/4* | — not offered | `browser_toolset_20260801` (client toolset, `browser_state` result blocks)
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].type`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*≈6,600-token definition; 7 models* | — not offered | browser is an `environment` of the computer-use tool, not a separate toolset
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` | no | Anthropic-only as a dedicated toolset.
*Ref:* tools.json |
## Tools — server side / hosted
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Web search**
*4/4* | `tools[type=web_search] {search_context_size, user_location, filters.allowed_domains, external_web_access, return_token_budget, search_content_types}`; `web_search_call` item + `url_citation` annotations; $10/1k calls
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=web_search]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Chat: search models + `web_search_options`; preview variants LEGACY* | `web_search_20260318\|20260209\|20250305 {max_uses, allowed_domains XOR blocked_domains, user_location}`; `server_tool_use` + `web_search_tool_result` blocks, `web_search_result_location` citations; $10/1k searches; dynamic filtering via code execution (20260209+)
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].max_uses`, `tools[].allowed_domains`, `tools[].blocked_domains`, `tools[].user_location`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*13 models; `pause_turn` after 10 iterations* | `tools[type=web_search] {allowed_domains ≤5 XOR excluded_domains ≤5, enable_image_understanding, enable_image_search}` → `web_search_call {action: search\|open_page\|find_in_page}` + `url_citation` annotations + inline `[[N]](url)` (disable via `include: no_inline_citations`); **$5/1k successful calls**; `search_context_size` → 400
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=web_search]`, `tools[].allowed_domains`, `tools[].excluded_domains`, `tools[].enable_image_search`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Responses only; Chat Live Search → 410; usage `server_side_tool_usage_details.web_search_calls`* | `tools[{googleSearch: {timeRangeFilter?, searchTypes?{webSearch, imageSearch}}}]` → `groundingMetadata {webSearchQueries, searchEntryPoint (must be displayed — ToS), groundingChunks[].web, groundingSupports}`; Gemini 3.x **5,000 free queries/month then $14/1k queries**, 2.5: 1,500 RPD free then $35/1k grounded prompts; legacy `googleSearchRetrieval` (dynamic retrieval) DEPRECATED
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].googleSearch`, `tools[].googleSearch.searchTypes`, `tools[].googleSearchRetrieval`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*429 `limit: 0` on this free-tier key; Interactions `{type: google_search}`; the only tool allowed with functions in the Live API* | yes | Same task on all four with four price models ($10 / $10 / $5 per 1k calls; Gemini per query with a free monthly quota). Domain filters exist on OpenAI, Anthropic and xAI; Gemini offers time-range and image-search filters instead.
*Ref:* docs/tools/openai/web-search.md · docs/tools/anthropic/web-search.md · docs/tools/xai/web-search.md · docs/tools/gemini/google-search-grounding.md |
| **Social / vertical search (X posts, Google Maps)**
*2/4* | — not offered | — not offered | `tools[type=x_search] {allowed_x_handles ≤10 XOR excluded_x_handles, from_date, to_date, enable_image_understanding, enable_video_understanding}` → live item `custom_tool_call` named `x_keyword_search`/`x_semantic_search` (docs: `x_search_call`) + `url_citation` to x.com posts; $5/1k calls until **2026-09-21**, then $5/1k posts + $10/1k profiles fetched
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=x_search]`, `tools[].allowed_x_handles`, `tools[].from_date`, `tools[].to_date`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` · `LIVE_DISCOVERED`
*also in the voice session* | `tools[{googleMaps: {enableWidget?}}]` + `toolConfig.retrievalConfig {latLng, languageCode}` → `groundingChunks[].maps {uri, title, placeId, text, placeAnswerSources}`; GA; text only; not in Live; Gemini 3.x $14/1k queries after 5,000/month (tools table: $25/1k grounded prompts, 1,500 RPD free)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].googleMaps`, `toolConfig.retrievalConfig.latLng`
**Status:** `DOCUMENTED` · `GA` · `LIVE_VERIFIED`
*pricing inconsistency flagged (`DOCUMENTATION_INCOMPLETE`)* | no | Provider-specific data sources; no equivalent on OpenAI/Anthropic.
*Ref:* docs/tools/xai/x-search.md · docs/tools/gemini/google-maps-grounding.md |
| **Web fetch / URL context (retrieve specific URLs)**
*2/4* | no dedicated fetch tool; web search may open pages; `input_file.file_url` fetches PDFs only
**Status:** `DOCUMENTED` | `web_fetch_20260318\|20260309\|20260209\|20250910 {max_uses, allowed_domains, citations, max_content_tokens, use_cache, url_sources}`; only URLs already in context; free (tokens only)
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].max_content_tokens`, `tools[].url_sources`, `tools[].use_cache`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`url_not_in_prior_context` error* | no fetch tool; `web_search_call.action.type: open_page\|find_in_page` shows the search tool opening pages; `input_file.file_url` attaches a remote file (attachment search)
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | `tools[{urlContext: {}}]` (no options): ≤20 public URLs per request, ≤34 MB each (HTML/JSON/text/CSV/RTF/PNG/JPEG/PDF; no YouTube/Workspace/paywalls) → `urlContextMetadata.urlMetadata[] {retrievedUrl, urlRetrievalStatus}`; free tool, content billed as input (`toolUsePromptTokenCount`)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].urlContext`
**Status:** `DOCUMENTED` · `GA` · `LIVE_VERIFIED`
*Interactions `{type: url_context}`; combinable with search/code exec/functions* | yes | Anthropic and Gemini offer a fetch tool; Anthropic restricts to URLs already in context, Gemini to any public URL you name.
*Ref:* docs/tools/anthropic/web-fetch.md · docs/tools/gemini/url-context.md |
| **File search / managed RAG**
*3/4* | Vector Stores API (create, files, file_batches, search) + `tools[type=file_search] {vector_store_ids, max_num_results, filters, ranking_options}`; $2.50/1k calls + $0.10/GB/day
**Endpoint:** `POST /v1/vector_stores · POST /v1/responses`
**Params:** `tools[type=file_search]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*16 endpoints* | no vector store; patterns: `search_result` blocks (your RAG, citable), `document` blocks from Files API, code execution over uploaded files
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].content[] {type:'search_result'}`, `messages[].content[] {type:'document'}`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | **Collections API** (`/v1/collections`, documents added from Files ids, `index_configuration.model_name: grok-embedding-small`, `chunk_configuration`, `POST /v1/documents/search {query, retrieval_mode: hybrid\|semantic\|keyword}`) + `tools[type=file_search\|collections_search] {vector_store_ids (= collection ids), max_num_results, filters, ranking_options}` → `file_search_call {queries, results[{file_id, filename, score, text}]}`; $2.50/1k calls + $0.10/GiB/day; implicit `attachment_search` over `input_file` parts $10/1k
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=file_search]`, `tools[].vector_store_ids`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` · `ACCOUNT_RESTRICTED`
*13 collection endpoints; documented on management-api.x.ai but working on api.x.ai (LIVE_DISCOVERED); search 404 while indexing* | **File Search stores** (`/v1beta/fileSearchStores`, `:uploadToFileSearchStore` resumable ≤100 MB, `:importFile`, documents, `chunkingConfig`, `customMetadata[]`; embedding model fixed at creation) + `tools[{fileSearch: {fileSearchStoreNames[], metadataFilter (AIP-160), topK}}]` → `groundingChunks[].retrievedContext {title, text, pageNumber, customMetadata}`; indexing $0.15/1M tokens once, storage + queries free; store quota 1 GB (free) … 1 TB (Tier 3)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].fileSearch`, `tools[].fileSearch.fileSearchStoreNames`, `tools[].fileSearch.metadataFilter`
**Status:** `DOCUMENTED` · `PREVIEW` · `LIVE_VERIFIED`
*12 endpoints; not combinable with Search/URL context; not in Live* | yes | Three hosted stores (OpenAI vector stores, xAI collections, Gemini file-search stores) with different billing (per call / per call + storage / per indexed token); Anthropic supplies citation plumbing instead of storage.
*Ref:* docs/openai/vector-stores.md · docs/anthropic/citations.md · docs/xai/collections.md · docs/tools/gemini/file-search.md |
| **Code execution (sandboxed Python)**
*4/4* | `tools[type=code_interpreter] {container: auto\|cntr_id, file_ids, memory_limit 1g–64g, network_policy}`; `code_interpreter_call {code, outputs}`; $0.03–$1.92 per 20-min session (per-minute since 2026-06-02)
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=code_interpreter].container`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*36 models* | `code_execution_20260521\|20260120\|20250825` (bash + text editor sub-tools, Python 3.11, 5 GiB RAM, no internet); `container {id, skills}` reuse (30-day state); 1,550 free container-hours/org/month then $0.05/h; free with web_search/fetch 20260209+
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].type`, `container`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*13 models; usage counter `code_execution_requests` missing live* | `tools[type=code_interpreter]` (alias `code_execution`) → `code_interpreter_call {code, outputs[{type: logs, logs: } \| {type: image, url}]}` (outputs only with `include: code_interpreter_call.outputs`); Python + NumPy/Pandas/Matplotlib/SciPy, no network; **$5/1k calls** + tokens; no container object
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=code_interpreter]`, `include`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*all 7 text models; gRPC rejects the `code_interpreter` alias* | `tools[{codeExecution: {}}]` → parts `executableCode {language: PYTHON, code}` + `codeExecutionResult {outcome, output}` (+ `inlineData` PNG for matplotlib); Python ≥3.10, fixed library set, 30 s per run, ≤5 retries, no pip/network; **no fee** — code and results billed as output then input tokens
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `tools[].codeExecution`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*12 models; Interactions `{type: code_execution}`; not in Live* | yes | Four sandboxes, four price models: per container-session (OpenAI), per container-hour with a free tier (Anthropic), per call (xAI), tokens only (Gemini). Only OpenAI/Anthropic expose a reusable container object.
*Ref:* docs/tools/openai/code-interpreter.md · docs/tools/anthropic/code-execution.md · docs/tools/xai/code-execution.md · docs/tools/gemini/code-execution.md |
| **Container management API**
*2/4* | CRUD containers and container files, download outputs
**Endpoint:** `GET/POST/DELETE /v1/containers[/{id}/files]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*9 endpoints; auto containers expire 20 min after last activity* | no container endpoints; container id returned on the message (`container.id`, `expires_at`) and reused via the `container` param; outputs via Files API
**Endpoint:** `POST /v1/messages · GET /v1/files/{id}/content`
**Params:** `container`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no container object; `container` param accepted for OpenAI compatibility, not needed
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | **Environments API** for Interactions agents (`POST/GET/DELETE /v1beta/environments`, files `GET …/files/{path}`, `PUT /upload/v1beta/environments/{env}/files/{path}`; sources repository/GCS/inline; network allowlist; idle 15 min, deleted after 7 days) — not usable by `codeExecution`
**Endpoint:** `POST /v1beta/environments`
**Params:** `environment`, `sources`, `network`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW` · `LIVE_VERIFIED`
*7 endpoints; sandbox compute unbilled during preview* | yes | OpenAI containers and Gemini environments are both first-class resources but serve different tools (code interpreter vs managed agents).
*Ref:* docs/openai/containers.md · docs/tools/anthropic/code-execution.md · docs/gemini/interactions-api.md |
| **Image generation as a tool / modality inside a text call**
*3/4* | `tools[type=image_generation] {model, quality, size, background, input_fidelity, partial_images, moderation}`; `image_generation_call` item; streaming partials
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=image_generation]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*31 models* | — not offered | `tools[type=image_generation]` → `image_generation_call` billed at Imagine per-image rates (no call fee); accepted live, not invoked
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=image_generation]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*all 7 text models listed* | not a tool: `generationConfig.responseModalities: [TEXT, IMAGE]` + `imageConfig {aspectRatio, imageSize 512\|1K\|2K\|4K}` on image models (`gemini-3.1-flash-image`, `-lite-image`, `gemini-3-pro-image`); Interactions `response_format {type: image}`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.responseModalities`, `generationConfig.imageConfig`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*no free tier (`limit: 0` here); text models ignore IMAGE modality* | yes | Tool item on OpenAI/xAI, output modality on Gemini; Claude outputs text only.
*Ref:* tools.json · docs/gemini/image-generation.md · docs/xai/images.md |
| **Remote MCP servers**
*4/4* | `tools[type=mcp] {server_label, server_url\|connector_id\|tunnel_id, authorization, headers, allowed_tools, require_approval (default always), defer_loading, allowed_callers}`; `mcp_list_tools`, `mcp_call`, `mcp_approval_request/response` items; no beta header
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=mcp]`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*42 models; also Realtime and Agents API* | `mcp_servers[] {type:url, url, name, authorization_token}` (≤20) + `tools[type=mcp_toolset] {mcp_server_name, default_config, configs}`; `mcp_tool_use`/`mcp_tool_result` blocks; **beta header `mcp-client-2025-11-20`**
**Endpoint:** `POST /v1/messages`
**Params:** `mcp_servers`, `tools[].mcp_server_name`, `tools[].default_config`, `tools[].configs`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*no approvals in Messages; the error text advertises `mcp-client-2026-09-15` which is rejected (inconsistency)* | `tools[type=mcp] {server_url, server_label (required), server_description, allowed_tools[], authorization, headers}` → `mcp_call {server_label, name, arguments, output, error}`; tokens only; **no approval round-trip** (`require_approval` silently accepted); also inside the voice session
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=mcp]`, `tools[].server_url`, `tools[].server_label`, `tools[].allowed_tools`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*LIVE_VERIFIED with DeepWiki; `connector_id` unsupported* | server-side `tools[].mcpServers[] {name, streamableHttpTransport {url, headers, timeout}}` in the discovery schema (UNVERIFIED, no guide) and Interactions `{type: mcp_server, name, url, headers, allowed_tools}` (documented; not on Gemini 3 per the overview); SDK-side MCP (`mcpToTool()`, Python `ClientSession` in `tools`) runs the calls in **your** process (BETA)
**Endpoint:** `POST /v1beta/interactions`
**Params:** `tools[].mcpServers`, `tools[](mcp_server)`
**Status:** `DOCUMENTED` · `DOCUMENTATION_INCOMPLETE` · `UNVERIFIED`
*Streamable HTTP only; server names must not contain '-'* | yes | OpenAI GA with approvals; xAI GA without approvals; Anthropic beta without approvals; Gemini server-side MCP is documented but unverified — its verified path is SDK-side execution.
*Ref:* docs/tools/openai/mcp-and-connectors.md · docs/tools/anthropic/mcp-connector.md · docs/tools/xai/mcp.md · docs/tools/gemini/mcp.md |
| **Built-in SaaS connectors**
*1/4* | `connector_id`: dropbox, gmail, googlecalendar, googledrive, microsoftteams, outlookcalendar, outlookemail, sharepoint (+ OAuth `authorization`) — **deprecated** for models after 2025-09-01
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=mcp].connector_id`
**Status:** `DOCUMENTED` · `DEPRECATED` | — not offered | `connector_id` documented as unsupported
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | — not offered | no | OpenAI-only (deprecated).
*Ref:* docs/tools/openai/mcp-and-connectors.md |
| **Private-network MCP (tunnels) / agent credentials**
*3/4* | Secure MCP Tunnel: outbound `tunnel-client`, `tools[type=mcp].tunnel_id` (`tunnel_[a-z0-9]{32}`)
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=mcp].tunnel_id`
**Status:** `DOCUMENTED` | MCP tunnels API (`/v1/tunnels`, certificates, tokens) + tunnel agent; header `mcp-tunnels-2026-06-22`; WIF bearer with `workspace:manage_tunnels`
**Endpoint:** `GET/POST /v1/tunnels`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW` · `ACCOUNT_RESTRICTED`
*10 endpoints; older `/v1/organizations/tunnels` deprecated* | — not offered | no tunnels; **Credentials API** (`/v1beta/credentials`: `bearer_token`, `oauth2`, `environment_variable` with `injection_location` and `trusted_domains`; secrets write-only) referenced from environment network allowlists
**Endpoint:** `POST /v1beta/credentials`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW`
*5 endpoints, not tested* | no | Tunnels on OpenAI/Anthropic; Gemini solves the secret-injection half with credentials; nothing on xAI.
*Ref:* endpoints.json (managed-agents tunnels, gemini credentials) · docs/tools/openai/mcp-and-connectors.md |
| **Skills (packaged instructions + files)**
*3/4* | Skills API (`POST /v1/skills`, versions, content zip) referenced from `tools[type=shell].environment.skills[]`, containers, Agents sessions; ≤500 files, ≤25 MB
**Endpoint:** `POST /v1/skills`
**Params:** `tools[type=shell].environment.skills`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` · `FAILED_VERIFICATION`
*version-by-number endpoints returned 404 live; lists returned empty* | Skills API (`/v1/skills`, versions, content; GA no header) + Anthropic skills (pptx, xlsx, docx, pdf); used via `container.skills[]` with code execution and on Managed Agents
**Endpoint:** `POST /v1/skills`
**Params:** `container.skills`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*9 endpoints all LIVE_VERIFIED* | Skills API `GET/POST /v1/skills`, `GET/DELETE /v1/skills/{id}`, `GET …/content` (zip; `SKILL.md` frontmatter `name`, `description`, `when-to-use`, `paths`, `allowed-tools`) — **all 404 for this team** (ACCOUNT_RESTRICTED); reach inference only via `shell.environment.skills[]`; Grok Build CLI also loads skills locally
**Endpoint:** `POST /v1/skills`
**Params:** `tools[type=shell].environment.skills`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*5 endpoints, OpenAPI only* | no Skills API; custom Interactions agents mount `.agents/skills//SKILL.md` from their environment sources
**Endpoint:** `POST /v1beta/interactions`
**Status:** `DOCUMENTED` · `PREVIEW` | yes | Same `SKILL.md` bundle model on OpenAI, Anthropic and xAI (xAI gated); Gemini uses environment files.
*Ref:* docs/openai/skills-api.md · docs/anthropic/skills-api.md · docs/xai/skills-api.md · docs/gemini/interactions-api.md |
| **Advisor (consult a stronger model mid-response)**
*1/4* | — not offered | `advisor_20260301 {model, max_tokens, caching}` server tool, beta `advisor-tool-2026-03-01`
**Endpoint:** `POST /v1/messages`
**Params:** `tools[].model`, `tools[].max_tokens`, `tools[].caching`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*11 models* | — not offered | — not offered | no | Anthropic-only.
*Ref:* tools.json |
| **Citations / grounding metadata**
*4/4* | `output_text.annotations[]`: `url_citation`, `file_citation`, `container_file_citation`, `file_path` (produced by web/file search, code interpreter)
**Endpoint:** `POST /v1/responses`
**Params:** `include`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no citations for plain documents* | `citations {enabled:true}` on `document`/`search_result` blocks → `char_location`, `page_location`, `content_block_location`, `search_result_location`, `web_search_result_location`; `citations_delta` when streaming
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].content[].citations`, `tools[].citations`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*all-or-nothing across documents* | `output_text.annotations[] {type: url_citation, url, start_index, end_index, title}` (live indices 0/0, title = URL) + inline markdown `[[N]](url)`; collections citations `collections:///files/`; `include: web_search_call.action.sources`
**Endpoint:** `POST /v1/responses`
**Params:** `include`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`usage.num_sources_used`* | `candidates[].groundingMetadata {groundingChunks[] (web\|maps\|retrievedContext), groundingSupports[] {segment, groundingChunkIndices, confidenceScores}, webSearchQueries, searchEntryPoint}`, `urlContextMetadata`, `citationMetadata` (recitation); Interactions `text_annotation_delta`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `candidates[].groundingMetadata`, `candidates[].urlContextMetadata`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*streaming chunks carry only new grounding chunks — accumulate* | yes | Anthropic cites any document you pass; OpenAI, xAI and Gemini cite only their own tool results (Gemini with segment-level supports and confidence scores).
*Ref:* docs/anthropic/citations.md · docs/tools/gemini/google-search-grounding.md |
## Multimodal input & media APIs
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Image input**
*4/4* | `input_image {image_url\|file_id, detail: low\|high\|auto\|original}`; patch/tile token formulas; ≤1,500 images, ≤512 MB
**Endpoint:** `POST /v1/responses`
**Params:** `input[](message).content[](input_image)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `image {source: base64\|url\|file, transformations}`; ≤100 (200k models) / 600 (1M models) images; 10 MB each; tokens = ⌈w/28⌉×⌈h/28⌉ (2576 px hi-res on 4.7+)
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].content[] {type:'image'}`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | Responses `input_image {image_url (https or data URL), detail}` (≥512 px; `file_id` variant → 400); Chat `image_url {url, detail}`; image tokens priced as text input (`image_input` = `input` price); all Grok 4 text models accept images
**Endpoint:** `POST /v1/responses`
**Params:** `input[](message).content[](input_image)`, `messages[].content[](image_url)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`usage.prompt_tokens_details.image_tokens`* | `parts[].inlineData {mimeType, data}` or `fileData {fileUri}` (Files API); `mediaResolution` LOW/MEDIUM/HIGH globally or per part (Gemini 3 adds ULTRA_HIGH, 2,240 tokens); docs 280/560/1,120/2,240 tokens by level (observed 256/529/1,089/2,209)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].parts[].inlineData`, `contents[].parts[].fileData`, `generationConfig.mediaResolution`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`promptTokensDetails[] {modality: IMAGE}`* | yes | Universal. Token accounting differs on every provider (patches/tiles, 28-px grid, text-price tokens, resolution levels).
*Ref:* docs/openai/multimodal-input.md · docs/anthropic/vision-and-documents.md · docs/xai/chat-completions.md · docs/gemini/multimodal-input.md |
| **PDF / document input**
*4/4* | `input_file {file_id\|file_url\|file_data+filename, detail}`; text + page images in context; <50 MB combined; office formats via file_id
**Endpoint:** `POST /v1/responses`
**Params:** `input[](message).content[](input_file)`
**Status:** `DOCUMENTED` | `document {source: base64 pdf\|url\|file\|text\|content, title, context, citations}`; ≤600 pages/request, 32 MB body
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].content[] {type:'document'}`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*citable* | Responses `input_file {file_id\|file_url\|file_data}` → routed through the implicit **attachment search** tool ($10/1k calls); Chat `file` parts → 400; `/v1/messages` `document` → 422
**Endpoint:** `POST /v1/responses`
**Params:** `input[](message).content[](input_file)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Files API 50 MB (spec) / 512 MB (guide)* | `inlineData {mimeType: application/pdf}` or `fileData` (Files API, 2 GB); PDF pages billed at the image token rate (`DOCUMENT` modality, 560 tokens/page in countTokens vs `IMAGE 520` observed in generateContent); `pdf_input` true on all 3.x text models
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].parts[].inlineData`, `contents[].parts[].fileData`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no citations for documents; URL context handles remote PDFs* | yes | Universal; only Anthropic makes documents citable, only xAI bills a per-call fee for attachments.
*Ref:* docs/anthropic/vision-and-documents.md · docs/xai/files.md · docs/gemini/multimodal-input.md |
| **Audio / video input (understanding)**
*2/4* | Chat Completions `input_audio {data, format}` on `gpt-audio-1.5`, `gpt-4o-audio-preview`; Responses `input_audio` part in schema but UNVERIFIED; no video input
**Endpoint:** `POST /v1/chat/completions`
**Params:** `messages[](user).content[](input_audio)`
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED` | — not offered | no audio/video parts on text models (audio only in the voice session and STT; `grok-imagine-video` accepts video/audio as generation inputs; `view_x_video` sub-tool understands X videos)
**Status:** `DOCUMENTED` | native: `inlineData`/`fileData` audio (WAV/MP3/AIFF/AAC/OGG/FLAC; ≈25–32 tokens/s) and video (≤1 fps frames + audio; `videoMetadata {startOffset, endOffset, fps}`; YouTube URLs via `fileData`) on every 3.x text model; `gemini-embedding-2` embeds audio/video too
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].parts[].inlineData`, `contents[].parts[].videoMetadata`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`audio_input` price rows on 2.5/3.1-lite/3-flash; 3.5+ single price* | yes | Gemini is the only provider with native audio **and** video understanding in the text API; OpenAI has audio-in on dedicated chat models.
*Ref:* docs/openai/multimodal-input.md §4 · docs/gemini/multimodal-input.md |
| **Text-to-speech**
*3/4* | `POST /v1/audio/speech` (`gpt-4o-mini-tts`, `tts-1`, `tts-1-hd`; SSE `speech.audio.delta`) and Chat `modalities:[text,audio]`
**Endpoint:** `POST /v1/audio/speech`
**Params:** `model`, `input`, `voice`, `instructions`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*custom voices ACCOUNT_RESTRICTED* | — not offered | `POST /v1/tts {text ≤60,000 chars, voice_id, language (required), output_format {codec mp3\|wav\|pcm\|mulaw\|alaw, sample_rate, bit_rate}, speed, with_timestamps}` + `wss://api.x.ai/v1/tts` streaming (`text.delta` → `audio.delta`); 28 built-in voices (`GET /v1/tts/voices`) + custom voices (`/v1/custom-voices`, Enterprise); **$15 / 1M characters**
**Endpoint:** `POST /v1/tts`
**Params:** `text`, `voice_id`, `language`, `output_format`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*17 voice endpoints; voice models absent from `GET /v1/models`* | `generateContent` on TTS models (`gemini-3.1-flash-tts-preview`, `gemini-2.5-flash\|pro-preview-tts`) with `responseModalities: [AUDIO]` + `speechConfig.voiceConfig.prebuiltVoiceConfig.voiceName` (30 voices) or `multiSpeakerVoiceConfig` (≤2 speakers); raw PCM 24 kHz out; $1 text in / $20 audio out per 1M (≈ $0.03/min); free tier (10 req/day observed)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.responseModalities`, `generationConfig.speechConfig`
**Status:** `DOCUMENTED` · `PREVIEW` · `LIVE_VERIFIED`
*streaming TTS on 3.1 only; multi-speaker may return `finishReason: OTHER`* | yes | Dedicated endpoint (OpenAI, xAI) vs a generation modality (Gemini); none on Anthropic.
*Ref:* docs/openai/audio.md · docs/xai/voice.md · docs/gemini/speech-generation.md |
| **Speech-to-text / transcription**
*3/4* | `POST /v1/audio/transcriptions` (`gpt-transcribe`, `gpt-4o-transcribe(-diarize)`, `whisper-1`; streaming `transcript.text.delta`), `POST /v1/audio/translations` (whisper-1)
**Endpoint:** `POST /v1/audio/transcriptions`
**Params:** `model`, `file`, `response_format`, `stream`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*whisper-1 & gpt-4o-transcribe shutdown 2027-02-26* | — not offered | `POST /v1/stt` (multipart `file` ≤500 MB or `url`; `language`, `diarize`, `keyterm`, `vad_threshold`, `model: grok-voice-transcribe-2.0\|1.0`) **$0.10/h**; `wss://api.x.ai/v1/stt` streaming (binary frames, `interim_results`, `endpointing`, `smart_turn` → `transcript.partial/done`) **$0.20/h**
**Endpoint:** `POST /v1/stt`
**Params:** `file`, `url`, `language`, `diarize`, `model`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*25 languages; default model contradicts between release notes (1.0) and model page (2.0)* | `gemini-3.5-transcribe` via `generateContent` + `generationConfig.audioTranscriptionConfig {languageCodes, customVocabulary, mode VERBATIM\|SMART, diarization, wordTimestamp}` → `parts[].audioTranscription {text, speakerLabel, words[]}` (LIVE_DISCOVERED shape); ≤1 h audio; ≈ $0.005/min blended; `gemini-3.5-transcribe-live` over the Live WebSocket (10-min sessions); `gemini-3.5-live-translate-preview` speech-to-speech translation
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.audioTranscriptionConfig`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*free tier available* | yes | OpenAI/xAI dedicated endpoints (per minute / per hour); Gemini a dedicated model behind the generic endpoint; none on Anthropic.
*Ref:* docs/openai/audio.md · docs/xai/voice.md · docs/gemini/transcription.md |
| **Realtime speech-to-speech (WebSocket / WebRTC / SIP)**
*3/4* | Realtime API GA: WebSocket `wss://api.openai.com/v1/realtime?model=`, WebRTC `POST /v1/realtime/calls`, SIP, ephemeral `POST /v1/realtime/client_secrets`; transcription & translation sessions; 60-min sessions
**Endpoint:** `WS /v1/realtime`
**Params:** `session.update`, `response.create`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*legacy `/v1/realtime/sessions` → 404* | — not offered | `wss://api.x.ai/v1/realtime?model=grok-voice-latest` (= `grok-voice-think-fast-2.0`) with the **OpenAI Realtime event vocabulary** (`session.update`, `input_audio_buffer.*`, `conversation.item.create`, `response.create` → `response.output_audio.delta`, `response.done`; 39 server / 9 client events); tools in-session (`function`, `web_search`, `x_search`, `file_search`, `mcp`); `reasoning.effort high\|none`; ephemeral `POST /v1/realtime/client_secrets` (≤3600 s); SIP via `/v2/phone-numbers`, `/v1/realtime/calls/{id}/refer\|hangup`, webhook `realtime.call.incoming`; **$0.08/min** + $0.004 per text item; 120-min sessions; concurrent sessions 10–200 by tier
**Endpoint:** `WS wss://api.x.ai/v1/realtime`
**Params:** `session.update`, `conversation.item.create`, `response.create`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*text turn LIVE_VERIFIED; audio UNVERIFIED; `ping` undocumented* | **Live API** `wss://generativelanguage.googleapis.com/ws/google.ai.generativelanguage.v1beta.GenerativeService.BidiGenerateContent` (`setup` → `setupComplete`; `clientContent`/`realtimeInput {audio\|video\|text\|activityStart\|activityEnd}`/`toolResponse` → `serverContent {modelTurn, turnComplete, interrupted, inputTranscription, outputTranscription}`, `toolCall`, `goAway`, `sessionResumptionUpdate`); models `gemini-3.8-live` (native audio, interleaved thinking, GA 2026-09-15), `-extended-thinking`, `gemini-3.1-flash-live-preview`, 2.5 native audio; ephemeral tokens `POST /v1beta/auth_tokens` on the `…Constrained` method; 16 kHz PCM in / 24 kHz out; ≈10-min connections, 15-min audio sessions (unlimited with `contextWindowCompression`), resumption handles 2 h; audio $3 in / $12 out per 1M (≈ $0.005 / $0.018 per min); free tier
**Endpoint:** `WSS BidiGenerateContent`
**Params:** `setup`, `realtimeInput`, `toolResponse`, `setup.sessionResumption`, `setup.contextWindowCompression`
**Status:** `DOCUMENTED` · `PREVIEW` · `LIVE_VERIFIED`
*32 message types recorded; `responseModalities: [TEXT]` → close 1007; only googleSearch + functions as tools* | yes | Three voice stacks: xAI copies OpenAI's Realtime protocol (so client code ports almost verbatim), Gemini's Live API is a distinct message family; Anthropic has none. Only OpenAI and xAI offer WebRTC/SIP.
*Ref:* docs/openai/realtime.md · docs/xai/voice.md · docs/gemini/live-api.md · docs/gemini/live-events.md |
| **Full-duplex voice with backend delegation (Live) / voice agents**
*1/4* | Live API: `POST /v1/live/sessions` (WebRTC), `WS /v1/live/sessions`, fork/attach/SIP; `gpt-live-1` $0.05/min; delegation to Responses or client
**Endpoint:** `POST /v1/live/sessions`
**Params:** `session.delegation`, `session.store`
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED` | — not offered | no separate delegation product; the realtime session itself runs server tools (web/x/file search, MCP) with `reasoning.effort`
**Endpoint:** `WS wss://api.x.ai/v1/realtime`
**Status:** `DOCUMENTED` | no delegation layer; `gemini-3.8-live-extended-thinking` runs background reasoning and speaks fillers (`interactionStatus: IN_PROGRESS`); `proactivity.proactiveAudio` always on for 3.8
**Endpoint:** `WSS BidiGenerateContent`
**Status:** `DOCUMENTED` | no | OpenAI-only as a product; xAI and Gemini fold agentic behaviour into the voice session itself.
*Ref:* docs/openai/live.md · docs/xai/voice.md · docs/gemini/live-api.md |
| **Image generation API**
*3/4* | `POST /v1/images/generations`, `/edits` (gpt-image-2, gpt-image-2.5-flare/sunburst; DALL·E RETIRED, gpt-image-1.x DEPRECATED); streaming partials; arbitrary sizes ≤4K
**Endpoint:** `POST /v1/images/generations`
**Params:** `prompt`, `size`, `quality`, `background`, `output_format`, `stream`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`/variations` RETIRED (404)* | — not offered | `POST /v1/images/generations {model, prompt, n 1–10, response_format url\|b64_json, aspect_ratio (incl. 21:9, 5:2, auto), resolution 1k\|1.5k\|2k, quality low\|medium\|auto, storage_options}` and `POST /v1/images/edits` (**JSON only**: `image {url\|file_id}` or `images[]` 2–5); `grok-imagine-image` $0.02, `grok-imagine-image-2.0` $0.04–$0.08, `grok-imagine-image-quality` $0.05 (DEPRECATED → 2026-11-02); sync ~5–10 s; always JPEG
**Endpoint:** `POST /v1/images/generations`
**Params:** `prompt`, `n`, `aspect_ratio`, `resolution`, `quality`, `response_format`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no `size`/`mask`/`seed`; `grok-2-image` RETIRED; not on us.api.x.ai* | `generateContent` on image models with `responseModalities` + `imageConfig {aspectRatio (14 ratios), imageSize 512\|1K\|2K\|4K}`; editing = pass the image as input and iterate; SynthID watermark; `gemini-3.1-flash-image` $0.045–$0.151, `-lite-image` $0.034, `gemini-3-pro-image` $0.134/$0.24; `gemini-2.5-flash-image` DEPRECATED → 2026-10-02; Imagen (`:predict`) RETIRED 2026-08-17; OpenAI-compat `POST /v1beta/openai/images/generations` subset
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.responseModalities`, `generationConfig.imageConfig.aspectRatio`, `generationConfig.imageConfig.imageSize`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*no free tier (`limit: 0`); batch 50 %* | yes | OpenAI and xAI have dedicated image endpoints (token-priced vs per-image); Gemini generates images through the text endpoint; Claude produces no images.
*Ref:* docs/openai/images.md · docs/xai/images.md · docs/gemini/image-generation.md |
| **Video generation API**
*3/4* | `/v1/videos` (Sora 2) — **DEPRECATED, shutdown 2026-09-24**; $0.10–$0.70/s
**Endpoint:** `POST /v1/videos`
**Params:** `prompt`, `seconds`, `size`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED` | — not offered | `POST /v1/videos/generations {model, prompt, duration 1–15, aspect_ratio, resolution 480p\|720p\|1080p, generate_audio, image, reference_images[], reference_audios[] (1.5), last_frame (1.5)}` → `{request_id}`; `GET /v1/videos/{request_id}` 202 pending → 200 `{status: done, video:{url}}`; `POST /v1/videos/edits`, `/extensions`; `grok-imagine-video` **$0.05/s**, `grok-imagine-video-1.5` **$0.08/s** (1080p, reference-to-video)
**Endpoint:** `POST /v1/videos/generations`
**Params:** `prompt`, `duration`, `aspect_ratio`, `resolution`, `image`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*4 endpoints; also via Batch (URLs expire 1 h)* | Veo 3.1: `POST /v1beta/models/veo-3.1-*:predictLongRunning {instances[{prompt, image, lastFrame, referenceImages[], video}], parameters {aspectRatio, resolution 720p\|1080p\|4k, durationSeconds 4\|6\|8, personGeneration, negativePrompt, seed}}` → `Operation`; poll `GET /v1beta/{name}`; download `files/{id}:download?alt=media`; native audio; $0.40/$0.60 (Veo 3.1), $0.10–$0.30 (Fast), $0.05/$0.08 (Lite) per second; `gemini-omni-1.1-flash` video-out model via Interactions (≈ $0.10/s); OpenAI-compat `POST /v1beta/openai/videos`
**Endpoint:** `POST /v1beta/models/{model}:predictLongRunning`
**Params:** `instances[].prompt`, `parameters.resolution`, `parameters.durationSeconds`
**Status:** `DOCUMENTED` · `PREVIEW`
*Veo 2.0/3.0 RETIRED 2026-06-30; no free tier; 400 on this key* | yes | Async polling everywhere; OpenAI's is shutting down, xAI's is the cheapest per second, Gemini's the only one with 4K.
*Ref:* docs/openai/video.md · docs/xai/videos.md · docs/gemini/video-generation.md |
| **Music generation**
*1/4* | — not offered | — not offered | — not offered | `lyria-3.5` (GA, **$0.08/song**), `lyria-3-clip-preview` ($0.04 / 30-s clip), `lyria-3-pro-preview` via `generateContent` (MP3) or Interactions `response_format {type: audio}` (WAV); **Lyria RealTime** WebSocket `…BidiGenerateMusic` (`weightedPrompts`, `musicGenerationConfig {bpm, density, brightness, scale…}`, `playbackControl`) → 48 kHz stereo PCM chunks (experimental, unpriced)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].parts[].text`, `musicGenerationConfig`, `playbackControl`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*no free tier for Lyria 3.x; RealTime LIVE_VERIFIED* | no | Gemini-only.
*Ref:* docs/gemini/music-generation.md |
| **Embeddings**
*2/4* | `text-embedding-3-small` ($0.02/M), `-3-large` ($0.13/M), `ada-002`; `dimensions`, `encoding_format`; 8,192 tokens/input, 2,048 inputs
**Endpoint:** `POST /v1/embeddings`
**Params:** `input`, `model`, `dimensions`, `encoding_format`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | — not offered | `POST /v1/embeddings {model, input, encoding_format, dimensions}` documented (OpenAPI) but `grok-embedding-small` → 404 'does not exist or your team does not have access', `GET /v1/embedding-models` → `{models: []}`; no published price; used internally to index Collections
**Endpoint:** `POST /v1/embeddings`
**Params:** `input`, `model`, `dimensions`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*gRPC `Embedder.Embed` exists* | `POST /v1beta/models/gemini-embedding-2:embedContent {content, outputDimensionality 128–3072, taskType (001 only), embedContentConfig}` and `:batchEmbedContents {requests[]}`; **multimodal** (text, ≤6 images, ≤180 s audio, ≤120 s video, PDF ≤6 pages → one vector); $0.20 text / $0.45 image / $6.50 audio / $12 video per 1M, batch 50 %; free tier; `gemini-embedding-001` DEPRECATED → 2028-05-14; OpenAI-compat `/v1beta/openai/embeddings`
**Endpoint:** `POST /v1beta/models/{model}:embedContent`
**Params:** `content`, `outputDimensionality`, `taskType`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`:asyncBatchEmbedContent` ACCOUNT_RESTRICTED (400 FAILED_PRECONDITION)* | yes | OpenAI (text) and Gemini (multimodal) ship embeddings; xAI's endpoint is documented but inaccessible; Anthropic has none.
*Ref:* docs/openai/embeddings.md · docs/gemini/embeddings.md · docs/models/xai-models.md |
| **Moderation endpoint**
*1/4* | `POST /v1/moderations` (`omni-moderation-latest`, text+image, free) and inline `moderation {model, policy}` on Responses/Chat
**Endpoint:** `POST /v1/moderations`
**Params:** `input`, `model`, `moderation`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`text-moderation-*` RETIRED* | — not offered | — not offered | no moderation route among the 125 Gemini endpoints; `safetySettings` thresholds and `safetyRatings` instead
**Status:** `DOCUMENTED` | no | OpenAI-only.
*Ref:* docs/openai/moderation.md · docs/gemini/safety.md |
| **Content provenance**
*1/4* | `POST /v1/content_provenance_checks` (C2PA verification)
**Endpoint:** `POST /v1/content_provenance_checks`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | generated media from code execution carry C2PA credentials; no verification endpoint
**Status:** `DOCUMENTED` | no provenance endpoint; `respect_moderation` flag on image/video results
**Status:** `DOCUMENTED` | SynthID watermark on all generated images/video/audio (Nano Banana 2 Lite also C2PA); no verification endpoint on the Developer API
**Status:** `DOCUMENTED` | no | Only OpenAI exposes a check endpoint.
*Ref:* endpoints.json · docs/gemini/image-generation.md |
## Files & batch
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Files API**
*4/4* | `POST /v1/files` (purposes user_data, batch, evals, assistants, vision, fine-tune), list/retrieve/delete/content; 512 MB/file; `expires_after` 1 h–30 d; batch files auto-expire 30 d
**Endpoint:** `POST /v1/files`
**Params:** `file`, `purpose`, `expires_after`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`user_data` content not downloadable (400)* | `POST /v1/files` (any MIME; 500 MB/file; workspace-scoped), list (`ids[]`), metadata, content (tool-generated files only), delete; `expires_in_seconds` 1 h–90 d; GA no header
**Endpoint:** `POST /v1/files`
**Params:** `file`, `expires_in_seconds`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*not on Bedrock/Vertex* | `POST /v1/files` (multipart; `purpose` ignored; `expires_after` 1 h–30 d), list (`filter` AIP-160, `sort_by`), get, delete, `GET …/content?format=original`, **public URLs** (`POST …/public-url`, `…/revoke`; ≤50 MiB, ≤1,000 active/team), chunked upload `files:initialize`/`:uploadChunks` (UNVERIFIED); storage $0.025/GiB/day, downloads $0.20/GiB; used by Responses `input_file`, Collections, Imagine, Batch
**Endpoint:** `POST /v1/files`
**Params:** `file`, `expires_after`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*10 endpoints; disabled under ZDR* | `POST /upload/v1beta/files` (resumable `X-Goog-Upload-*` protocol, multipart or media), `GET /v1beta/files[/{id}]`, `DELETE`, `POST /v1beta/files:register {uris: gs://…}` (GCS, 30 days), `GET /v1beta/generatedFiles`, download `files/{id}:download?alt=media` (generated files); **2 GB/file, 20 GB/project, 48 h TTL, free**; used via `fileData {fileUri}`
**Endpoint:** `POST /upload/v1beta/files`
**Params:** `file`, `displayName`, `mimeType`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*unknown file → 403 (never 404); not available on Vertex AI* | yes | All four are blob stores referenced by id; TTLs range from 48 h (Gemini, fixed) to 90 d (Anthropic); only xAI prices storage and only xAI mints public URLs.
*Ref:* docs/openai/files-and-uploads.md · docs/anthropic/files-api.md · docs/xai/files.md · docs/gemini/files.md |
| **Multipart / resumable upload**
*3/4* | Uploads API: ≤8 GB in ≤64 MB parts, 1 h TTL
**Endpoint:** `POST /v1/uploads`
**Params:** `filename`, `purpose`, `bytes`, `mime_type`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | — not offered | `POST /v1/files:initialize` + `POST /v1/files:uploadChunks` (documented, UNVERIFIED)
**Endpoint:** `POST /v1/files:initialize`
**Status:** `DOCUMENTED` · `UNVERIFIED` | Google resumable upload protocol on `/upload/v1beta/files` and `/upload/v1beta/fileSearchStores/{s}:uploadToFileSearchStore` (`start` → `x-goog-upload-url` → `upload, finalize`; 8 MiB chunk granularity); environments `PUT /upload/v1beta/environments/{env}/files/{path}` (≤2 GiB)
**Endpoint:** `POST /upload/v1beta/files`
**Params:** `X-Goog-Upload-Protocol`, `X-Goog-Upload-Command`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | Three resumable schemes, all provider-specific.
*Ref:* docs/openai/files-and-uploads.md · docs/xai/files.md · docs/gemini/files.md |
| **Batch processing (async, discounted)**
*4/4* | JSONL file (`custom_id`, `method`, `url`, `body`) → `POST /v1/batches {input_file_id, endpoint, completion_window:'24h'}`; 50,000 requests / 200 MB; endpoints responses, chat, embeddings, completions, moderations, images, videos; 50 % off; output file 30 d
**Endpoint:** `POST /v1/batches`
**Params:** `input_file_id`, `endpoint`, `completion_window`, `output_expires_after`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*cancel LIVE_DISCOVERED 409 on terminal* | inline `requests[] {custom_id, params}` → `POST /v1/messages/batches`; 100,000 requests / 256 MB; 24 h expiry; results JSONL at `results_url` 29 d; 50 % off all token dims (stacks with caching); all Messages features incl. server tools; `output-300k-2026-03-24` beta
**Endpoint:** `POST /v1/messages/batches`
**Params:** `requests`, `requests[].custom_id`, `requests[].params`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*6 endpoints all LIVE_VERIFIED* | `POST /v1/batches {name}` → `POST /v1/batches/{id}/requests {batch_requests[{batch_request_id, batch_request: {chat_get_completion \| responses \| image_generation \| image_edit \| video_generation \| video_extension}}]}` (inline, ≤25 MB/request) or JSONL via Files (`input_file_id`, ≤50,000 lines / 200 MB); `GET …/requests`, `GET …/results`, `POST …:cancel`; **20 % off** on grok-4.3 / 4.20 only (grok-4.6 / 4.5 / build unsupported; Imagine at standard rates); bypasses rate limits; not with priority; live turnaround 11 s
**Endpoint:** `POST /v1/batches`
**Params:** `name`, `input_file_id`, `batch_requests`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*8 endpoints; DELETE → 405; `responses` bodies come back as `chat_get_completion`* | `POST /v1beta/models/{model}:batchGenerateContent {batch:{displayName, inputConfig:{requests:{requests[]} \| fileName}, priority, webhookConfig}}` (inline ≤20 MB or JSONL Files ≤2 GB) → `Operation` `batches/{id}`; `:asyncBatchEmbedContent`; `GET /v1beta/batches[/{id}]`, `:cancel`, `DELETE`, `PATCH :update*Batch`; states PENDING→RUNNING→SUCCEEDED\|FAILED\|CANCELLED\|EXPIRED (>48 h); **50 %**; target 24 h; 100 concurrent jobs; per-model enqueued-token caps; results 6 weeks; OpenAI-compat `/v1beta/openai/batches`
**Endpoint:** `POST /v1beta/models/{model}:batchGenerateContent`
**Params:** `batch.inputConfig.requests`, `batch.inputConfig.fileName`, `batch.displayName`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*9 endpoints; free tier → 400 FAILED_PRECONDITION; list LIVE_VERIFIED* | yes | Same 24-h / discounted idea on all four, with 50 % (OpenAI, Anthropic, Gemini) vs 20 % (xAI, three models only). Inline requests: Anthropic, xAI, Gemini; file-based: OpenAI, xAI, Gemini.
*Ref:* docs/openai/batch.md · docs/anthropic/message-batches.md · docs/xai/batches.md · docs/gemini/batch.md |
## Prompt caching
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Prompt / context caching**
*4/4* | automatic prefix caching (≥1,024 tokens; exact prefix; `prompt_cache_key` routing hint); GPT-5.6+: `prompt_cache_options {mode: implicit\|explicit, ttl:'30m', prewarm}` and per-part `prompt_cache_breakpoint {mode:explicit}` (≤4); `prompt_cache_retention: in_memory\|24h` (deprecated)
**Endpoint:** `POST /v1/responses`
**Params:** `prompt_cache_key`, `prompt_cache_options`, `prompt_cache_retention`, `prompt_cache_breakpoint`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*reads 0.1× (5.6+) / model-specific; writes 1.25× only on GPT-5.6+* | explicit `cache_control {type: ephemeral, ttl: 5m\|1h}` on system/tools/content blocks (≤4 breakpoints) or top-level automatic; 20-block lookback; min 512 (Fable/Mythos/Opus 5) · 1,024 (Sonnet 5/4.6/4.5, Opus 4.8) · 2,048 (Opus 4.7) · 4,096 (Haiku 4.5, Opus 4.6/4.5)
**Endpoint:** `POST /v1/messages`
**Params:** `cache_control`, `system[].cache_control`, `messages[].content[].cache_control`, `tools[].cache_control`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*writes 1.25× (5m) / 2× (1h); reads 0.1× (0.025× Fable 5.1)* | **automatic** prefix cache, no markers (`cache_control` on `/v1/messages` ignored); sticky routing via `prompt_cache_key` (Responses, Chat) or header `x-grok-conv-id` (Chat/gRPC); include prior `reasoning_content` / encrypted reasoning or use `previous_response_id` to keep hits; no TTL or minimum documented, no write charge; cached reads 0.15×–0.25× (grok-4.6 $0.50, grok-4.3 $0.20); long-context tier doubles cached price too
**Endpoint:** `POST /v1/responses`
**Params:** `prompt_cache_key`, `x-grok-conv-id`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*live: a fresh minimal request already shows ~192 cached tokens (hidden system prefix); cached tokens count toward TPM* | **implicit** caching (automatic on 2.5+, min prompt 4,096 tokens on 3.x Flash/3.1 Pro, 2,048 on 2.5; hits in `usageMetadata.cachedContentTokenCount`) **and explicit** `cachedContents` (`POST /v1beta/cachedContents {model, contents, systemInstruction, tools, ttl\|expireTime}` → use `cachedContent: cachedContents/{id}`; default TTL 1 h, PATCH TTL only; min 1,024 tokens live); cached reads 0.1× everywhere; explicit storage $0.50–$4.50 per 1M tokens per hour
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `cachedContent`, `generationConfig`, `ttl`, `expireTime`
**Status:** `DOCUMENTED` · `BETA` · `ACCOUNT_RESTRICTED`
*explicit caching v1beta-only, not via Interactions; free tier `limit: 0`; no write fee* | yes | Two philosophies: implicit (OpenAI default, xAI, Gemini implicit) vs explicit breakpoints (Anthropic, OpenAI 5.6+, Gemini `cachedContents`). Write fees exist only on Anthropic and GPT-5.6+; Gemini charges storage time instead; xAI charges nothing but publishes no TTL.
*Ref:* docs/comparisons/caching-and-reasoning.md · docs/xai/prompt-caching.md · docs/gemini/context-caching.md |
| **Cache diagnostics**
*2/4* | `prompt_cache_options.comparison_response_id` → `prompt_cache_diagnostics {type: cache_hit\|cache_miss{reason}}` (GPT-5.6+)
**Endpoint:** `POST /v1/responses`
**Params:** `prompt_cache_options.comparison_response_id`
**Status:** `DOCUMENTED`
*reasons: model_changed, tools_changed, text_format_changed, …* | `diagnostics.previous_message_id` → response `diagnostics` cache-miss reasons; beta `cache-diagnosis-2026-04-07`
**Endpoint:** `POST /v1/messages`
**Params:** `diagnostics.previous_message_id`
**Status:** `DOCUMENTED` · `BETA` | no diagnostics; only `usage.*cached_tokens` counters (`input_tokens_details.cached_tokens`, `prompt_tokens_details.cached_tokens`, `cache_read_input_tokens` on `/v1/messages`)
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | no diagnostics; `usageMetadata.cachedContentTokenCount` + `cacheTokensDetails[]` only
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` | yes | OpenAI and Anthropic explain misses; xAI and Gemini only count hits.
*Ref:* parameters.json |
## Reasoning / thinking
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Reasoning control**
*4/4* | `reasoning {effort: none\|minimal\|low\|medium\|high\|xhigh\|max, summary: auto\|concise\|detailed, context: auto\|current_turn\|all_turns, mode: standard\|pro}`; `reasoning` output items with `encrypted_content`; `usage.output_tokens_details.reasoning_tokens`
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.effort`, `reasoning.summary`, `reasoning.context`, `reasoning.mode`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Chat: `reasoning_effort` only; effort set per model (gpt-6-astra rejects `none`)* | `thinking {type: adaptive\|enabled\|disabled, budget_tokens, display: summarized\|omitted\|updates}` + `output_config.effort: low\|medium\|high\|xhigh\|max` (default high); `thinking`/`redacted_thinking` blocks with `signature`; `usage.output_tokens_details.thinking_tokens`
**Endpoint:** `POST /v1/messages`
**Params:** `thinking.type`, `thinking.budget_tokens`, `thinking.display`, `output_config.effort`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`enabled` (manual budget) 400 on 4.7+; adaptive 400 on 4.5 models* | Responses `reasoning {effort: low\|medium\|high\|xhigh, summary: auto\|concise\|detailed}` (alias `reasoning_effort`); Chat `reasoning_effort` + `message.reasoning_content` (`delta.reasoning_content` when streaming); `usage.completion_tokens_details.reasoning_tokens`; **reasoning is always on** for grok-4.6/4.5/4.20-reasoning/build (`reasoning_effort` → 400 on 4.20-reasoning and build); grok-4.3 accepts `none` (LIVE_DISCOVERED, 0 reasoning tokens); `grok-4.20-0309-non-reasoning` has none; on `grok-4.20-multi-agent-0309` effort selects **4 or 16 agents**
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.effort`, `reasoning.summary`, `reasoning_effort`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`max_output_tokens` documented to include reasoning but not enforced live; 'Reply with OK.' costs 70–180 reasoning tokens* | `generationConfig.thinkingConfig {thinkingLevel: MINIMAL\|LOW\|MEDIUM\|HIGH (Gemini 3), thinkingBudget: -1\|0\|N (2.5-era, LEGACY on 3.x; exclusive with level → 400), includeThoughts}`; defaults high (3.1 Pro, 3 Flash) / medium (3.5–3.8 Flash) / minimal (Flash-Lite); `MINIMAL` → 400 on 3.8/3.7 Flash and 3.1 Pro; thought parts `{text, thought: true}`; `usageMetadata.thoughtsTokenCount`; Interactions `generation_config.thinking_level` + `thinking_summaries: auto\|none`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.thinkingConfig.thinkingLevel`, `generationConfig.thinkingConfig.thinkingBudget`, `generationConfig.thinkingConfig.includeThoughts`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*thinking cannot be disabled on 3.8/3.7 Flash, 3.1 Pro, 2.5 Pro; `thinkingBudget: 0` works on 3.6/3.5 Flash; OpenAI-compat maps `reasoning_effort` minimal/low/medium/high* | yes | Four effort ladders overlap on low/medium/high: OpenAI adds none/minimal/xhigh/max, Anthropic xhigh/max, xAI xhigh, Gemini minimal. Off-switches: OpenAI `none` (most models), Anthropic `disabled` (not Fable/Mythos), xAI `none` on grok-4.3 only, Gemini `thinkingBudget: 0` on two Flash models only. Only Anthropic and Gemini (2.5) expose a token budget.
*Ref:* docs/openai/reasoning.md · docs/anthropic/thinking.md · docs/xai/reasoning.md · docs/gemini/thinking.md |
| **Reasoning visibility**
*4/4* | summaries only (`reasoning.summary`); raw `content[]` empty for GPT models; encrypted replay item
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.summary`, `include`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | summarized text by default (≤4.6) or omitted (4.7+); `display: updates` progress text (Fable 5.x beta); signature always present
**Endpoint:** `POST /v1/messages`
**Params:** `thinking.display`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | Chat returns the **full `reasoning_content` text** by default; Responses returns `reasoning` items with `summary[]` (`reasoning.summary` accepted 'for compatibility'; the item is sometimes omitted); `/v1/messages` returns `thinking` blocks with an empty `signature`
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.summary`, `include`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `includeThoughts: true` → thought **summaries** as `thought: true` parts (not guaranteed on trivial prompts); `thoughtSignature` (opaque) on function-call and final parts; Interactions `thought` steps `{signature, summary[]}`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.thinkingConfig.includeThoughts`, `contents[].parts[].thoughtSignature`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*billing is on full thoughts, not the summary* | yes | Only xAI (Chat Completions) exposes the raw chain of thought; the other three return summaries plus an opaque signature/encrypted blob.
*Ref:* docs/anthropic/thinking.md §3 · docs/xai/reasoning.md · docs/gemini/thinking.md |
| **Reasoning replay across turns**
*4/4* | replay `reasoning` items verbatim (auto with `previous_response_id`); `encrypted_content` decrypted in memory when `store:false`; `reasoning.context: all_turns` (GPT-5.6 default)
**Endpoint:** `POST /v1/responses`
**Params:** `input[](reasoning).encrypted_content`, `reasoning.context`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | pass `thinking` blocks back unmodified within tool-use turns (400 if modified); prior-turn thinking kept for all turns on Opus 4.5+/Sonnet 4.6+/Fable, last turn on Haiku/Sonnet 4.5; `clear_thinking_20251015` edit; beta `thinking.block_binding`
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].content[].signature`, `thinking.block_binding`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `include: ["reasoning.encrypted_content"]` (xai-sdk `use_encrypted_content=True`) → replay the `reasoning` items with `encrypted_content` when `store:false` (ZDR) — also the 'top cause of cache misses' when omitted; `previous_response_id` rehydrates automatically; Chat: resend `reasoning_content`
**Endpoint:** `POST /v1/responses`
**Params:** `include`, `input[](reasoning).encrypted_content`, `previous_response_id`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | **mandatory** on Gemini 3: echo `thoughtSignature` of the first `functionCall` of each step (400 'missing a thought_signature' / `finishReason: MISSING_THOUGHT_SIGNATURE`, even at `minimal`); text-part signatures recommended; tampering → 400 'Corrupted thought signature'; Interactions carry it in `thought` steps; 'thought preservation' across turns since 3.5 Flash
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `contents[].parts[].thoughtSignature`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*dummy values `skip_thought_signature_validator` bypass validation (verified)* | yes | All four carry hidden reasoning as an opaque blob; Anthropic and Gemini validate it (400 on edits / omissions), OpenAI and xAI encrypt it.
*Ref:* docs/anthropic/thinking.md §4 · docs/xai/reasoning.md · docs/gemini/thinking.md |
| **Pro / extended-compute mode**
*3/4* | `reasoning.mode: pro` (GPT-5.6) billed at standard rates; `*-pro` models (gpt-5.4-pro, gpt-5.5-pro)
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.mode`
**Status:** `DOCUMENTED` | `effort: max` is the top of the ladder; no separate mode
**Params:** `output_config.effort`
**Status:** `DOCUMENTED` | `grok-4.20-multi-agent-0309` (BETA): one Responses call fans out to 4 (low/medium) or 16 (high/xhigh) agents; all agents' tokens billed at grok-4.20 rates; Responses-only (Chat → 400)
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.effort`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*'Reply with OK.' cost 1,645 output tokens ≈ $0.0047* | `gemini-3.8-live-extended-thinking` (Live) and Deep Research **agents** (`deep-research-preview-04-2026`, `-max-`; 60-min background runs, $1–7 per task estimated) rather than a mode; `gemini-3.1-pro-preview` is the deep-reasoning text model
**Endpoint:** `POST /v1beta/interactions`
**Params:** `agent`, `agent_config(deep-research)`
**Status:** `DOCUMENTED` · `PREVIEW` · `LIVE_VERIFIED` | no | Different vehicles: a mode/model (OpenAI), a multi-agent model (xAI), an agent (Gemini), a ladder value (Anthropic).
*Ref:* parameters.json · docs/models/xai-models.md · docs/gemini/interactions-api.md |
| **Change effort mid-conversation without breaking the cache**
*2/4* | `configuration_update` input item `{reasoning:{effort}}` (gpt-6-astra)
**Endpoint:** `POST /v1/responses`
**Params:** `input[](configuration_update)`
**Status:** `DOCUMENTED` | `messages[].output_config.effort` on `role:system` entries; beta `mid-conversation-output-config-2026-07-01` (Fable 5.1, Mythos 5.1, Opus 5)
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].output_config.effort`
**Status:** `DOCUMENTED` · `BETA` | per-request `reasoning.effort` only (automatic cache; no documented invalidation rule)
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | per-request `thinkingLevel` only; an explicit `cachedContents` prefix is unaffected by generationConfig changes
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` | yes | OpenAI and Anthropic only.
*Ref:* parameters.json |
| **Task-wide token budget (advisory)**
*3/4* | none in Responses (`max_tool_calls` caps hosted tool calls; Agents API has `session_budget_exceeded`)
**Endpoint:** `POST /v1/responses`
**Params:** `max_tool_calls`
**Status:** `DOCUMENTED` | `output_config.task_budget {type, total ≥20000, remaining}`; beta `task-budgets-2026-03-13` (Fable/Mythos/Opus 5/4.8/4.7)
**Endpoint:** `POST /v1/messages`
**Params:** `output_config.task_budget`
**Status:** `DOCUMENTED` · `BETA` | `max_turns` (agentic turns per Responses request; default = server cap) — a turn budget, not tokens
**Endpoint:** `POST /v1/responses`
**Params:** `max_turns`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | Antigravity `agent_config.max_total_tokens` (Interactions, PREVIEW) → status `incomplete` when exhausted
**Endpoint:** `POST /v1beta/interactions`
**Params:** `agent_config(antigravity).max_total_tokens`
**Status:** `DOCUMENTED` · `PREVIEW` | no | Three different budget units (tokens, turns, agent tokens); none on the OpenAI Responses API.
*Ref:* compatibility/anthropic-feature-model-matrix.json · docs/xai/responses.md · docs/gemini/interactions-api.md |
## Context management
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Context window**
*4/4* | 1,050,000 (GPT-5.4/5.5/5.6/6 Astra; >272k input = long-context pricing 2×/1.5×), 400k (5.x mini/nano, GPT-5–5.3), 200k (o-series), 128k (gpt-4o)
**Status:** `DOCUMENTED` | 1,000,000 default on Claude 4.6+ (no header, no premium); 200k on Opus 4.5 / Sonnet 4.5 / Haiku 4.5; `context-1m-2025-08-07` header RETIRED
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | 1,000,000 (grok-4.3, all grok-4.20 ids), 500,000 (grok-4.6, grok-4.5), 256,000 (grok-build-0.1); prompts ≥200,000 tokens switch the **whole request** to the 2× long-context tier
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`usage.context_details` on Responses* | 1,048,576 on every Gemini 3.x / 2.5 text model (`inputTokenLimit`); 262,144 Gemma 4; 131,072 Live / image / agent models; Pro models bill >200k prompts at 2× input / 1.5× output, Flash models flat
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | ~1M on every flagship line; long-context premiums on OpenAI (>272k), xAI (≥200k, all tokens) and Gemini Pro (>200k); none on Anthropic or Gemini Flash.
*Ref:* models.json |
| **Max output tokens**
*4/4* | 128,000 on GPT-5.x/6 (272,000 on gpt-5-pro); `max_output_tokens` ≥16
**Endpoint:** `POST /v1/responses`
**Params:** `max_output_tokens`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | 128,000 on Claude 4.6+ (64,000 on 4.5 models); `max_tokens` required; 300,000 in Batches with `output-300k-2026-03-24`
**Endpoint:** `POST /v1/messages`
**Params:** `max_tokens`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | **no documented limit** (`max_output: null` on every Grok record; grok-4.6 page: 'No text output limit'); Responses `max_output_tokens` default 128,000 (docs) and not enforced on reasoning live; Chat `max_completion_tokens` caps visible output only
**Endpoint:** `POST /v1/responses`
**Params:** `max_output_tokens`, `max_completion_tokens`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`max_tokens` DEPRECATED alias on Chat* | 65,536 on every 3.x / 2.5 text model (`outputTokenLimit`); 32,768 Gemma 4 / image models; `maxOutputTokens` **includes thinking tokens** (hit → `finishReason: MAX_TOKENS`, possibly empty text)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `generationConfig.maxOutputTokens`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Interactions `generation_config.max_output_tokens`* | yes | 128k (OpenAI, Anthropic), 64k (Gemini), unpublished (xAI). Only Anthropic makes the cap mandatory.
*Ref:* models.json |
| **Server-side compaction (in-flight)**
*3/4* | `context_management: [{type: compaction, compact_threshold ≥1000}]` → `compaction` output item, SSE `response.compaction.compacting`
**Endpoint:** `POST /v1/responses`
**Params:** `context_management[].compact_threshold`
**Status:** `DOCUMENTED` | `context_management.edits: [{type: compact_20260112, trigger ≥50000, pause_after_compaction, instructions}]` → `compaction` block, `stop_reason: compaction`; beta `compact-2026-01-12`; 4.6+
**Endpoint:** `POST /v1/messages`
**Params:** `context_management.edits[]`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*usage.iterations[] for billing* | `context_management[]` accepted but 'parsed but not yet executed' (compat only); use the stand-alone compact endpoint
**Endpoint:** `POST /v1/responses`
**Params:** `context_management`
**Status:** `DOCUMENTED` | Live API only: `setup.contextWindowCompression {triggerTokens, slidingWindow {targetTokens}}` (unlimited session length); Antigravity agents compact around ~135k automatically; nothing on `generateContent`
**Endpoint:** `WSS BidiGenerateContent`
**Params:** `setup.contextWindowCompression`
**Status:** `DOCUMENTED` | yes | OpenAI and Anthropic compact text conversations in-flight; Gemini only compresses Live sessions; xAI parses the field and ignores it.
*Ref:* docs/openai/responses.md §6 · docs/anthropic/context-management.md §3 · docs/gemini/live-api.md |
| **Stand-alone compaction request**
*3/4* | `POST /v1/responses/compact {model, input\|previous_response_id}` → `response.compaction` with encrypted `compaction` item
**Endpoint:** `POST /v1/responses/compact`
**Params:** `model`, `input`, `previous_response_id`, `instructions`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `compaction: {type: summarize}` on Messages → single signed `compaction` block; beta `compact-2026-09-04`
**Endpoint:** `POST /v1/messages`
**Params:** `compaction`, `compaction.type`, `compaction.instructions`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*block must be sent first* | `POST /v1/responses/compact {model, input}` → `{object: response.compaction, id: cmp_…, output:[{type: compaction, encrypted_content}], usage {…, dropped_message_count}}`; put the `output` first in the next `input`; do not edit the blob; the pre-compaction conversation must still fit the window (May 2026 'Context Compaction API')
**Endpoint:** `POST /v1/responses/compact`
**Params:** `model`, `input`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*48 parameter rows* | — not offered | yes | OpenAI-shaped on xAI (same endpoint and item); Anthropic as a Messages parameter; none on Gemini.
*Ref:* docs/anthropic/context-management.md §3 · docs/xai/responses.md |
| **Server-side context editing (clear old tool results / thinking)**
*1/4* | no editing strategies; legacy `truncation: auto` drops oldest items
**Endpoint:** `POST /v1/responses`
**Params:** `truncation`
**Status:** `DOCUMENTED` · `LEGACY` · `LIVE_VERIFIED` | `context_management.edits[]`: `clear_tool_uses_20250919`, `clear_thinking_20251015`; beta `context-management-2025-06-27`; response `context_management.applied_edits[]`
**Endpoint:** `POST /v1/messages`
**Params:** `context_management.edits[].type`, `context_management.edits[].trigger`, `context_management.edits[].keep`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `truncation` accepted (`disabled` echoed) — 'not supported, compatibility only'
**Endpoint:** `POST /v1/responses`
**Params:** `truncation`
**Status:** `DOCUMENTED` | — not offered | no | Anthropic-only.
*Ref:* docs/anthropic/context-management.md §2 |
| **Mid-conversation tool add/remove (cache-preserving)**
*2/4* | `additional_tools` developer item (adds tools mid-thread); `tool_choice: allowed_tools` to restrict
**Endpoint:** `POST /v1/responses`
**Params:** `input[](additional_tools)`
**Status:** `DOCUMENTED` | `tool_addition`/`tool_removal` blocks in `role:system` messages; beta `mid-conversation-tool-changes-2026-07-01` (Fable 5, Mythos 5, Opus 4.8, Opus 5)
**Endpoint:** `POST /v1/messages`
**Params:** `messages[].role`
**Status:** `DOCUMENTED` · `BETA` | follow-ups via `previous_response_id` 'may change tools/model' — no dedicated item; cache effect undocumented
**Endpoint:** `POST /v1/responses`
**Params:** `previous_response_id`
**Status:** `DOCUMENTED` | resend the full `tools[]` each call (Interactions chaining also requires re-sending tools)
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` | yes | OpenAI and Anthropic only.
*Ref:* anthropic-beta-headers.json |
## Service tiers, limits, safety
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Service tiers / processing modes**
*4/4* | `service_tier: auto\|default\|flex\|scale\|priority\|fast\|ultrafast`; flex = batch price, slower; fast = 2× (renamed from priority 2026-07-30); echoed in response
**Endpoint:** `POST /v1/responses`
**Params:** `service_tier`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `service_tier: auto\|standard_only` (Priority Tier commitments, no longer sold) + `speed: fast` (beta `fast-mode-2026-02-01`, Opus 5 / 4.8 only, 2× price); `usage.service_tier` standard\|priority\|batch
**Endpoint:** `POST /v1/messages`
**Params:** `service_tier`, `speed`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW` · `ACCOUNT_RESTRICTED`
*fast 429 'rate limit of 0' for our key* | `service_tier: default\|priority` (Chat + Responses): **priority = 2×** on every token type (after the cache discount), billed only when the response echoes `service_tier: priority`; not combinable with Batch
**Endpoint:** `POST /v1/responses`
**Params:** `service_tier`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*live responses always echoed `default`* | `serviceTier: standard\|flex\|priority` (`service_tier` on Interactions / OpenAI-compat): **flex = 0.5×** (1–15 min target, sheddable → 429), **priority = 1.8×** (0.3× rate limit, graceful downgrade to standard); echoed in `usageMetadata.serviceTier` and header `X-Gemini-Service-Tier`
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `serviceTier`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*flex accepted live; ten text models listed* | yes | Cheaper tier: OpenAI flex, Gemini flex (both 50 %). Faster tier: OpenAI fast 2×, Anthropic fast 2× (two models), xAI priority 2×, Gemini priority 1.8×.
*Ref:* docs/anthropic/service-tiers.md · docs/xai/pricing.md · docs/gemini/pricing.md · pricing.json |
| **End-user identifier for abuse detection**
*4/4* | `safety_identifier` (≤64 chars; replaces `user`) + `prompt_cache_key`; header `OpenAI-Safety-Identifier` on Realtime
**Endpoint:** `POST /v1/responses`
**Params:** `safety_identifier`, `user`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `metadata.user_id` (≤512 chars, no PII); beta `anthropic-user-profile-id` header
**Endpoint:** `POST /v1/messages`
**Params:** `metadata.user_id`, `anthropic-user-profile-id`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `user` and `safety_identifier` (both accepted on Chat and Responses); `/v1/messages` `metadata.user_id`
**Endpoint:** `POST /v1/responses`
**Params:** `safety_identifier`, `user`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `labels {safety_identifier: …}` (documented key; Cloud-label rules; accepted, not echoed) on generateContent and Interactions
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `labels`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | Same idea everywhere; the field name changes four times.
*Ref:* parameters.json |
| **Rate-limit headers**
*3/4* | `x-ratelimit-{limit,remaining,reset}-{requests,tokens}` (+ `-project-tokens`), `Retry-After`, `retry-after-ms`, `x-should-retry`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*reset as Go durations (`6m0s`)* | `anthropic-ratelimit-{requests,tokens,input-tokens,output-tokens}-{limit,remaining,reset}` (RFC 3339), `retry-after`, `x-should-retry`, `anthropic-priority-*`, `anthropic-fast-*`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*not on GET /models or count_tokens* | **undocumented but observed**: `x-ratelimit-limit-requests` (7,200 grok-4.6/4.5, 1,800 grok-4.3/4.20/build — per minute), `x-ratelimit-remaining-requests`, `x-ratelimit-limit-tokens` (50,000,000 / 10,000,000 = documented T0 TPM), `x-ratelimit-remaining-tokens`; no `reset`, no `Retry-After`; `x-request-id`, `x-zero-data-retention`
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED`
*absent on `/v1/responses` multi-agent calls and catalogue GETs* | **none** — no `x-ratelimit-*` or `Retry-After` on 200/404/429; retry delay only inside the 429 message text ('Please retry in 54.2s') and `google.rpc.RetryInfo`; undocumented `X-Gemini-Service-Tier` header
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED`
*no request-id header either — use `responseId` in the body* | yes | Headers on three providers (xAI's undocumented); Gemini puts the information in the 429 body.
*Ref:* headers.json · rate-limits.json |
| **Rate-limit tiers**
*4/4* | Free, Tier 1–5 by cumulative spend ($5 → $1,000) with monthly usage caps $100 → $200k; per-model RPM/TPM/batch-queue tables; long-context tables >272k
**Status:** `DOCUMENTED`
*Scale/Reserved Tier, Ultrafast preview above Tier 5* | Start ($500/mo cap) / Build ($1,000) / Scale ($200k) / Custom; per model-class RPM/ITPM/OTPM (cache reads excluded from ITPM); Batches & count_tokens separate
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED`
*observed Scale-tier headers for our key* | Tier 0–4 by cumulative spend since 2026-01-01 ($0 / $50 / $250 / $1,000 / $5,000; Enterprise on request); per-model **RPS** (= RPM/60) and **TPM** (prompt + completion + reasoning + cached): grok-4.6/4.5 150→500 RPS, 50M→100M TPM; grok-4.3/4.20/build 37→208 RPS, 10M→85M TPM; multi-agent 9→56 RPS; Imagine RPS-only (6→100 images, 10→158 videos); voice concurrent sessions 10→200; Batch bypasses limits; per-key `qps/qpm/tpm` caps via Management API
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED`
*console shows personalised limits* | Free / Tier 1 (billing linked; $250 cap, $10 per rolling 10 min) / Tier 2 ($100 paid + 3 days; $2,000; $50) / Tier 3 ($1,000 + 30 days; $20k–100k+; $200); dimensions RPM / TPM / RPD (+ IPM, TPD) **per project**; per-model matrix published only in AI Studio; Pro and media models unavailable on Free (`limit: 0`); priority 0.3× limits; batch enqueued-token caps per model
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED`
*free-tier RPM observed 15/min on flash-lite* | yes | Spend-based tiers everywhere; xAI is the only provider publishing exact per-model RPS/TPM per tier in the docs, Gemini the only one with a genuinely free tier.
*Ref:* generated/rate-limits.json |
| **Overload / capacity error**
*4/4* | HTTP 503 `server_is_overloaded` (retryable, `Retry-After`)
**Status:** `DOCUMENTED` | HTTP 529 `overloaded_error` (also as SSE `error` event after 200); acceleration limits now 429
**Status:** `DOCUMENTED` | HTTP 429 (RPS/TPM/credits; gRPC RESOURCE_EXHAUSTED) and 5xx `internal` — no dedicated overload code; status.x.ai
**Status:** `DOCUMENTED`
*no 429 triggered in this run* | HTTP 503 `UNAVAILABLE` ('The model is overloaded. Please try again later.'), 429 `RESOURCE_EXHAUSTED` when Flex capacity is shed, 504 `DEADLINE_EXCEEDED` for long Flex/Deep Research requests
**Status:** `DOCUMENTED` | yes | 503 (OpenAI, Gemini), 529 (Anthropic), 429 (xAI) for the same condition.
*Ref:* errors.json |
| **Error envelope**
*4/4* | `{error: {message, type, param, code}}`; types invalid_request_error, rate_limit_error, insufficient_quota, server_error…; codes e.g. `model_not_found`, `context_length_exceeded`, `previous_response_not_found`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*empty-body 404 from Cloudflare on unknown URLs* | `{type: error, error: {type, message}, request_id}`; types invalid_request_error, authentication_error, permission_error, not_found_error, request_too_large, rate_limit_error, api_error, overloaded_error, billing_error, timeout_error
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*no `code` field except `error.details.error_code` on some 429/529* | `{code: , error: }` (`invalid-argument`, `not-found`, `unauthenticated:no-credentials`…); 422 = **bare JSON string** (serde message); Management API = gRPC-style `{code: 16, message, details[]}`; Realtime WS `{type: error, error:{type, code, message}}`; **an invalid API key returns 400, not 401**
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*12 error records; gRPC↔HTTP mapping 3→400, 16→401, 7→403, 5→404, 8→429* | google.rpc `Status`: `{error: {code: , message, status: INVALID_ARGUMENT\|FAILED_PRECONDITION\|UNAUTHENTICATED\|PERMISSION_DENIED\|NOT_FOUND\|ALREADY_EXISTS\|RESOURCE_EXHAUSTED\|INTERNAL\|UNIMPLEMENTED\|UNAVAILABLE\|DEADLINE_EXCEEDED…, details[] (BadRequest.fieldViolations, QuotaFailure, RetryInfo, Help)}}`; Interactions `{error: {code: , message}}`; soft failures at HTTP 200 (`promptFeedback.blockReason`, `finishReason`); Live WS close codes 1007/1008
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*18 error records; unknown File/Operation → 403, not 404* | yes | Four envelopes; only Anthropic and OpenAI carry a request id in the body/header pair; Gemini has no request-id header at all.
*Ref:* docs/errors/openai.md · docs/errors/anthropic.md · docs/xai/authentication-headers-errors.md · docs/errors/gemini.md |
| **Idempotency key**
*0/4* | not documented for api.openai.com (only Workspace Agents on api.chatgpt.com); dedupe via `metadata`/`custom_id`
**Status:** `UNVERIFIED` | not documented; SDKs retry on 409
**Status:** `UNVERIFIED` | not documented; `public-url` creation is idempotent by design; batch `batch_request_id` dedupes within a batch
**Status:** `UNVERIFIED` | not documented; batch `key` per JSONL line; `seed` for determinism
**Status:** `UNVERIFIED` | no | No provider offers request idempotency keys on the model APIs.
*Ref:* headers.json |
| **Per-request dollar cost in the response**
*1/4* | token counts only (`usage`); costs via the Admin `GET /v1/organization/costs` report
**Status:** `DOCUMENTED` | token counts only; costs via `/v1/organizations/cost_report`
**Status:** `DOCUMENTED` | `usage.cost_in_usd_ticks` on **every** inference response (Chat, Responses, images, videos; 1 USD = 10^10 ticks; April 2026) — the effective price after cache, long-context, priority and regional multipliers; batch `cost_breakdown` (SDK/gRPC)
**Endpoint:** `POST /v1/responses`
**Params:** `usage.cost_in_usd_ticks`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*catalogue endpoints return no usage/cost* | token counts only (`usageMetadata`, Interactions `usage`); costs in Google Cloud Billing
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `DOCUMENTED` | no | xAI-only.
*Ref:* docs/xai/pricing.md · docs/xai/responses.md |
## Auth, versioning, SDKs, platform
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Authentication**
*4/4* | `Authorization: Bearer `; optional `OpenAI-Organization`, `OpenAI-Project`; Admin keys `sk-admin-…`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*WIF token exchange at auth.openai.com / mTLS* | `x-api-key: sk-ant-api03-…` **or** `Authorization: Bearer `; `anthropic-workspace-id`; Admin keys `sk-ant-admin01-…`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*WIF via `POST /v1/oauth/token`* | `Authorization: Bearer xai-…` on REST, WebSocket and gRPC metadata; separate **Management key** for `management-api.x.ai` (inference key → 401 code 16); ephemeral `xai-realtime…` client secrets for browsers (`POST /v1/realtime/client_secrets`); per-key ACLs `api-key:endpoint:*`, `api-key:model:*` (new keys have **no access by default**); mTLS host `mtls.api.x.ai` (enterprise)
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`GET /v1/api-key`, `GET /v1/me` introspection* | `x-goog-api-key: AIza…` (recommended; `?key=` discouraged); `Authorization: Bearer ` **mandatory** on `/v1beta/openai/*`; OAuth/ADC Bearer alternative (+ `x-goog-user-project`); ephemeral tokens `POST /v1beta/auth_tokens` for the Live API; **standard keys rejected from September 2026** in favour of service-account-bound auth keys
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*limits are per project, not per key* | yes | Bearer everywhere except Gemini's native header; only OpenAI/Anthropic/xAI split admin vs inference keys; Gemini is the only one retiring a key type.
*Ref:* headers.json · docs/openai/authentication-and-keys.md · docs/anthropic/admin-api.md §1 · docs/xai/authentication-headers-errors.md · docs/gemini/authentication-headers-versions.md |
| **API version header / path version**
*2/4* | none (server answers `openai-version: 2020-10-01`)
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `anthropic-version: 2023-06-01` **required** on every request (400 otherwise)
**Params:** `anthropic-version`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no version header, no beta headers; features selected by body fields or base URL; Interactions-style `Api-Revision` does not exist
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | version in the **URL path**: `/v1beta` (86 methods, default for SDKs) vs `/v1` (47; stable subset — no caching, tuning, Live, Files, agents); Interactions optional header `Api-Revision: 2026-05-20`
**Params:** `Api-Revision`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*docs say every model is in both versions; live `/v1/models` lists 22 vs 58* | no | Header (Anthropic), path (Gemini), nothing (OpenAI, xAI).
*Ref:* headers.json · docs/gemini/authentication-headers-versions.md |
| **Beta opt-in header**
*2/4* | `OpenAI-Beta`: `agents=v1`, `chatkit_beta=v1`, `workspace_agent_runs=v1`, `responses_multi_agent=v1`, legacy `assistants=v2`, `realtime=v1`; plus `?beta=true` surface with body `openai-beta[]`
**Params:** `OpenAI-Beta`
**Status:** `DOCUMENTED`
*400 `invalid_beta` when missing* | `anthropic-beta: -[,…]` (50 catalogued values; SDK `betas=[…]`, `client.beta.*`); unknown → 400
**Params:** `anthropic-beta`, `betas`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*see docs/faq.md for the current list* | **none**; alpha features answer 403 ('only available for alpha users' — `tool_search`) or 404 (ACL)
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | **none**; gating by `/v1beta` path and `-preview` / `-exp` model ids (`kind: preview` 45 records)
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no | Anthropic gates parameters, OpenAI gates surfaces, Gemini gates by path/model id, xAI by account.
*Ref:* generated/fragments/headers/anthropic-beta-headers.json · headers.json · docs/faq.md Q6 |
| **Request correlation**
*3/4* | response `x-request-id`; request `X-Client-Request-Id` (logged, not echoed); `openai-processing-ms`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | response `request-id` (also `request_id` in error body); `anthropic-organization-id`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | response `x-request-id` (= `chat.completion.id`), `Server-Timing` (Cloudflare `cfEdge`/`cfOrigin`), `CF-RAY`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*absent on catalogue GETs* | no request-id header; `responseId` in the JSON body; `Server-Timing: gfet4t7; dur=…`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | Header on three providers, body field on Gemini.
*Ref:* headers.json |
| **Official SDKs**
*4/4* | Python `openai` 3.16.2, Node `openai` 7.18/7.19, .NET, Java 4.65 (beta label), Go v3 (beta), Ruby, CLI, Agents SDK (py/ts), Azure libraries; retries 2, timeout 600 s
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | Python `anthropic` 1.7.0, TS `@anthropic-ai/sdk` 0.126, Go, Java 2.63, Ruby, C# ≥10, PHP (beta), `ant` CLI 1.33; retries 2, timeout 10 min; cloud clients Bedrock/Vertex/AWS/Foundry
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*Python v1 removed sampling kwargs and `completions`* | Python `xai-sdk` 1.19 (**gRPC**, `client.chat.create().sample()/stream()/defer()/parse()`, `response.cost_usd`, timeout 1620 s) — no official Node SDK: use `openai` with `baseURL: https://api.x.ai/v1`, `@ai-sdk/xai`, `langchain-xai`, or the Anthropic SDK against `/v1/messages` (deprecated); Grok Build CLI (`grok`, BETA); protos `xai-org/xai-proto` for `buf curl`; gRPC `api.x.ai:443` (11 services / 38 RPCs) lacks Responses items, voice, skills
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*9 SDK records* | Python `google-genai` 2.24 (`genai.Client()`; same SDK targets Vertex with `vertexai=True`), JS `@google/genai` 2.23, Go `google.golang.org/genai`, Java `com.google.genai`, C# `Google.GenAI`; Firebase AI Logic / Genkit / Vercel AI SDK; OpenAI SDK against `/v1beta/openai`; legacy `google-generativeai`, `@google/generative-ai`, Go/Dart/Swift/Android libs **DEPRECATED** 2025-11-30
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*8 SDK records; default `api_version` v1beta* | yes | Every provider has Python and Node coverage — xAI's Node path is the OpenAI SDK; xAI's own SDK is the only gRPC-first one.
*Ref:* sdks.json · docs/xai/sdks.md · docs/gemini/sdks.md |
| **Webhooks (platform events)**
*4/4* | `/v1/webhook_endpoints` CRUD + `rotate_secret`, `test`; `/v1/webhook_event_types`; events batch.*, response.*, fine_tuning.*, eval.run.*, video.*, realtime/live incoming calls, safety.*, agent.session.*; Standard Webhooks signature (`webhook-id/-timestamp/-signature`, `whsec_`)
**Endpoint:** `GET /v1/webhook_endpoints`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*28 event types* | Managed Agents only: endpoints registered in Console (no API), 44 event types (agent.*, session.*, deployment*.*, environment.*, vault*.*, memory_store.*); same Standard Webhooks headers; ≤3 attempts, 5-min freshness; beta `managed-agents-2026-04-01`
**Status:** `DOCUMENTED` · `BETA`
*no webhooks for Messages/Batches* | only the SIP voice webhook `realtime.call.incoming` (Standard Webhooks headers, HMAC-SHA256) registered via `POST /v2/phone-numbers {webhook:{url}}`; no batch/response webhooks
**Endpoint:** `POST /v2/phone-numbers`
**Status:** `DOCUMENTED` | `POST/GET /v1/webhooks`, `GET\|PATCH\|DELETE /v1/webhooks/{id}`, `POST …/rotate_secret` (documented on **/v1**, BETA) + per-request `webhook_config {uris[], user_metadata}` / Batch `webhookConfig`; events `interaction.completed\|failed\|cancelled\|requires_action`, `batch.succeeded\|failed`; JWKS-signed
**Endpoint:** `POST /v1/webhooks`
**Params:** `webhook_config`
**Status:** `DOCUMENTED` · `BETA`
*7 endpoints, not tested; `:ping` UNVERIFIED* | yes | OpenAI covers the widest event set; Gemini and Anthropic cover their agent/interaction resources; xAI only incoming calls.
*Ref:* generated/webhook-events.json · docs/openai/webhooks.md · docs/anthropic/managed-agents.md §17.2 · docs/gemini/interactions-api.md |
| **Administration API**
*3/4* | 124 operations under `/v1/organization/*` and `/v1/projects/*`: admin keys, users, invites, projects, service accounts, project API keys, groups, roles, certificates (mTLS), data retention, spend limits/alerts, audit logs, usage & costs; Admin API key `sk-admin-…`
**Endpoint:** `GET /v1/organization/*`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*all probes 403/401 with a project key* | 100 operations under `/v1/organizations/*`: me, users, invites, workspaces, api_keys, service accounts, federation (WIF), RBAC, spend limits, external keys, tunnels certs, usage_report & cost_report, analytics, Claude Code analytics; Admin key `sk-ant-admin01-…`
**Endpoint:** `GET /v1/organizations/*`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` · `LIVE_VERIFIED`
*`GET /v1/organizations/me` works with a regular key* | **Management API** `https://management-api.x.ai` (21 endpoints, separate Management key, camelCase): API keys (create with `acls[]`, `qps/qpm/tpm`, `expireTime`; rotate; delete; propagation), team models/endpoints ACLs, billing (`billing-info`, invoices, payment methods, postpaid spending limits, prepaid balance/top-up, `POST …/usage` aggregated by key/model/IP/cluster), audit events; teams/members/ZDR toggle are console-only
**Endpoint:** `GET /auth/teams/{teamId}/api-keys`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED`
*inference key → 401 code 16* | no admin API on the Developer API: keys/projects/quota live in AI Studio and Google Cloud IAM/Console; `GET /v1beta/models` is the only account-scoped read
**Status:** `DOCUMENTED` | yes | Three admin surfaces (OpenAI, Anthropic, xAI — all inaccessible to this atlas's keys); Gemini delegates to Google Cloud.
*Ref:* docs/openai/admin-api.md · docs/anthropic/admin-api.md · docs/xai/management-api.md |
| **Usage & cost reporting**
*4/4* | `GET /v1/organization/usage/{completions,embeddings,images,…}` + `GET /v1/organization/costs` (buckets 1m/1h/1d, group_by)
**Endpoint:** `GET /v1/organization/costs`
**Params:** `start_time`, `bucket_width`, `group_by`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | `GET /v1/organizations/usage_report/messages`, `/usage_report/claude_code`, `/cost_report`; analytics `/analytics/{usage_report,cost_report,user_*}`
**Endpoint:** `GET /v1/organizations/cost_report`
**Params:** `starting_at`, `bucket_width`, `group_by`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | per-request **`usage.cost_in_usd_ticks`** (1 USD = 10^10 ticks) on every inference response + `POST /v1/billing/teams/{team_id}/usage` (Management API, aggregated by API key / model / IP / cluster / token type); batch `cost_breakdown` (SDK/gRPC)
**Endpoint:** `POST /v1/billing/teams/{team_id}/usage`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` · `LIVE_VERIFIED`
*cost ticks LIVE_VERIFIED; billing endpoint restricted* | no reporting endpoint; per-response `usageMetadata` (modality breakdown, `thoughtsTokenCount`, `toolUsePromptTokenCount`, `cachedContentTokenCount`, `serviceTier`) and Interactions `usage {total_*_tokens, *_by_modality, grounding_tool_count}`; billing dashboards in Google Cloud
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Params:** `usageMetadata`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | yes | Org-level reports on OpenAI/Anthropic/xAI; xAI is the only one returning the dollar cost per call; Gemini only per-response token detail.
*Ref:* endpoints.json (admin, management) · docs/xai/pricing.md · docs/gemini/generate-content.md |
| **Audit / compliance data access**
*3/4* | `GET /v1/organization/audit_logs` (scope `api.audit_logs.read`); safety alerts/cases API
**Endpoint:** `GET /v1/organization/audit_logs`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | Compliance API (36 endpoints, Claude Enterprise Compliance Access Key): activities feed, chats, projects, files, sessions, roles, groups; DELETE operations
**Endpoint:** `GET /v1/compliance/activities`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | `GET /audit/teams/{teamId}/events` (Management API; administrative events only)
**Endpoint:** `GET /audit/teams/{teamId}/events`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | none on the Developer API (Cloud Audit Logs on Vertex AI)
**Status:** `DOCUMENTED` | yes | Admin-action logs on OpenAI/Anthropic/xAI; Anthropic additionally exports end-user content (Claude Enterprise).
*Ref:* docs/anthropic/compliance-and-iam.md · docs/xai/management-api.md |
| **Spend limits**
*3/4* | org/project `spend_limit`, `spend_alerts` Admin endpoints; 429 `insufficient_quota` codes when hit
**Endpoint:** `POST /v1/organization/spend_limit`
**Status:** `DOCUMENTED` | `/v1/organizations/spend_limits`, `/spend_limits/effective`, increase requests approve/deny; tier cap → 429 without `retry-after`; self-set limit → 400
**Endpoint:** `POST /v1/organizations/spend_limits`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | `GET/POST /v1/billing/teams/{team_id}/postpaid/spending-limits`, prepaid balance / top-up (Management API); per-key `qps/qpm/tpm` caps
**Endpoint:** `POST /v1/billing/teams/{team_id}/postpaid/spending-limits`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | tier spend caps ($250 / $2,000 / $20k+) and rolling 10-minute spend limits ($10 / $50 / $200) are enforced by Google, not configurable via API (Cloud Billing budgets instead)
**Status:** `DOCUMENTED` | yes | API-settable on three providers; platform-imposed on Gemini.
*Ref:* endpoints.json (admin) · rate-limits.json (gemini) |
| **Cloud availability**
*4/4* | Azure OpenAI / Microsoft Foundry (`AzureOpenAI` clients, Azure libraries); regional hosts `us.\|eu.\|au.\|jp.\|in.api.openai.com`
**Status:** `DOCUMENTED`
*Azure surface not catalogued in this atlas* | Amazon Bedrock (Mantle `bedrock-mantle.{region}.api.aws/anthropic/v1/messages` and legacy InvokeModel), Google Cloud Vertex AI (`rawPredict`), Microsoft Foundry (`/anthropic/v1/*`), Claude Platform on AWS; feature gaps per platform (no Batches/Files/Skills/server tools on Bedrock/Vertex)
**Status:** `DOCUMENTED` · `UNVERIFIED`
*24 cloud endpoint records* | Google Cloud **Vertex AI Model Garden** (partner model) and **Microsoft Foundry** (Azure) — OpenAI-compatible chat + Responses, cloud billing; also OpenRouter, Vercel AI Gateway, Cloudflare, Cursor; **regional first-party endpoints** `https://us.api.x.ai/v1` (US-pinned, grok-4.6, 1.1×) and `https://eu-west-1.api.x.ai/v1` (undocumented, grok-4.3, LIVE_DISCOVERED); clusters us-east-1, us-west-2, us-central-1, eu-west-1, us-saltlake-2
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED` | Gemini Developer API (`generativelanguage.googleapis.com`, global, API key) **vs Vertex AI / Gemini Enterprise Agent Platform** (`{location}-aiplatform.googleapis.com`, IAM only, ~40 regions, data residency, ZDR, CMEK, VPC-SC, provisioned throughput, supervised tuning, RAG/Agent Engine); one SDK, two backends; Developer-API-only: Interactions, Live, Files, File Search, agents, free tier
**Status:** `DOCUMENTED`
*Vertex surface not catalogued in this atlas* | yes | Claude and Grok are resold on other clouds; OpenAI and Gemini have first-party cloud twins (Azure/Foundry, Vertex).
*Ref:* docs/anthropic/cloud-providers.md · docs/openai/data-residency-and-regions.md · docs/xai/sdks.md · docs/gemini/vertex-vs-gemini-api.md |
| **Data residency**
*3/4* | project region at creation; regional hosts `us./eu./au./jp./in.api.openai.com`; 10 % uplift for models ≥2026-03-05; EU lacks `background:true`
**Status:** `DOCUMENTED` · `ACCOUNT_RESTRICTED` | `inference_geo: us\|global` per request (Claude 4.6+; 1.1× multiplier for `us`); workspace `default_inference_geo`; `usage.inference_geo` echo; Bedrock/Vertex regional endpoints +10 %
**Endpoint:** `POST /v1/messages`
**Params:** `inference_geo`
**Status:** `DOCUMENTED` · `FAILED_VERIFICATION`
*400 on Haiku 4.5 live* | base-URL choice: `https://us.api.x.ai/v1` keeps request handling, inference, moderation and retained data in the US at **1.1×** token prices (grok-4.6 only; no image/video/voice); global `api.x.ai` gives no region guarantee
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*`eu-west-1.api.x.ai` serves grok-4.3 at global prices (undocumented)* | no residency option on the Developer API (global endpoint); residency, CMEK and ~40 regions are Vertex AI features
**Status:** `DOCUMENTED` | yes | Regional pinning costs ~10 % on OpenAI, Anthropic and xAI alike; Gemini requires moving to Vertex.
*Ref:* docs/anthropic/regions-and-ips.md · docs/openai/data-residency-and-regions.md · docs/xai/pricing.md · docs/gemini/vertex-vs-gemini-api.md |
| **Zero data retention / data-use terms**
*4/4* | ZDR/Modified retention by approval; `store:false` keeps reasoning encrypted; Live recordings and Agents sessions excluded
**Params:** `store`
**Status:** `DOCUMENTED` | Messages stateless → ZDR-eligible on Claude API (per model flag `zero_data_retention_eligible`); not eligible: Files API, MCP connector, programmatic tool calling, Managed Agents, Fable 5.1 / Mythos 5.1
**Status:** `DOCUMENTED` | team-wide **ZDR toggle** (console): response header `x-zero-data-retention: true\|false`, `GET /v1/me` → `zdr_status: no_zdr\|zdr`; disables `store`, `previous_response_id`, Files, Collections, Batch, deferred completions, stored media; default retention 30 days, not used for training; encrypted reasoning replay keeps ZDR + caching
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*our team: `no_zdr`* | **Unpaid Services** (free tier): prompts/responses may be used to improve Google products, with human review; **Paid Services**: not used for training, 55-day abuse logging; ZDR **not achievable** on the Developer API (grounding stores 30 days; Interactions state unless `store:false`) — Vertex AI offers ZDR; EEA/UK/CH end-user apps must use Paid Services
**Params:** `store`
**Status:** `DOCUMENTED` | yes | ZDR is a program (OpenAI), a per-model eligibility (Anthropic), a team switch (xAI) and unavailable on Gemini's Developer API.
*Ref:* models.json (anthropic capabilities.zero_data_retention_eligible) · docs/xai/management-api.md · docs/gemini/authentication-headers-versions.md |
| **OpenAI-compatibility layer**
*3/4* | the native surface (Responses + Chat Completions)
**Endpoint:** `POST /v1/chat/completions`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | none — Anthropic exposes only its own Messages format (partners such as xAI implement it)
**Status:** `DOCUMENTED` | **native**: the whole inference API is OpenAI-shaped (`/v1/chat/completions`, `/v1/responses` incl. items and SSE events, `/v1/batches`, `/v1/files`, `/v1/images/*`, `/v1/realtime` events); rejected/ignored OpenAI fields: `background`, `metadata` (Responses 400), `logit_bias` 400, `stop`/penalties on reasoning models 400, `logprobs` ignored, `store` ignored on Chat, non-function tools 422, `images.edit()` multipart unsupported; extras: `reasoning_effort`, `reasoning_content`, `deferred`, `max_turns`, `top_k`, `min_p`, `cost_in_usd_ticks`
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED`
*plus an Anthropic-compatible `/v1/messages` (deprecated)* | `/v1beta/openai/*` (BETA): `chat/completions` (LIVE_VERIFIED), `embeddings`, `models[/{id}]`, `images/generations` (subset), `videos` (Sora-style Veo), `batches`; Bearer key required; `reasoning_effort` → `thinkingLevel` mapping; `extra_body.google {thinking_config, cached_content, safety_settings, tools[{google_search}]}`; `message.extra_content.google.thought_signature`; unknown params silently ignored; **no** Responses, Assistants, audio, files, fine-tuning, moderations
**Endpoint:** `POST /v1beta/openai/chat/completions`
**Params:** `extra_body.google.thinking_config`, `extra_body.google.cached_content`, `reasoning_effort`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*7 endpoints; google.rpc error envelope* | yes | xAI is OpenAI-shaped by design; Gemini offers a partial adapter; Anthropic none.
*Ref:* docs/xai/chat-completions.md · docs/xai/responses.md · docs/gemini/openai-compatibility.md |
## Managed agents platforms
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Managed agent harness**
*4/4* | Agents API (beta `OpenAI-Beta: agents=v1`): `POST /v1/agents` (model, instructions, reasoning, text, tools, multi_agent), sessions, turns, items, events, environments, templates, vaults; Codex harness; 34 + 9 endpoints
**Endpoint:** `POST /v1/agents`
**Params:** `model`, `instructions`, `tools`, `multi_agent`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*gpt-5.4-nano rejected; gpt-5.6-luna accepted* | Claude Managed Agents (beta `managed-agents-2026-04-01`): `POST /v1/agents` (name, model, system, tools, mcp_servers, skills, multiagent) with **versions**, `/v1/sessions` + events/stream, environments, deployments (cron), vaults, memory stores, dreams, tunnels, user profiles; 96 endpoints
**Endpoint:** `POST /v1/agents`
**Params:** `name`, `model`, `system`, `tools`, `mcp_servers`, `skills`, `multiagent`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*Claude 4.5+ models* | no agent resource: the **Responses API agentic loop** (server-side tools iterate inside one request, bounded by `max_turns`; stored responses + `previous_response_id` = the session); `grok-4.20-multi-agent-0309` (BETA) as a built-in multi-agent model; **Grok Build** (`grok-build-0.1` PREVIEW model + `grok` CLI BETA: TUI/headless/ACP, MCP, hooks, skills, subagents, sandbox) is a client-side harness
**Endpoint:** `POST /v1/responses`
**Params:** `tools`, `max_turns`, `store`, `previous_response_id`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` · `BETA`
*Skills API 404 for this team* | **Interactions API** agents (`agent` instead of `model`): Deep Research (`deep-research-preview-04-2026`, `-max-`, `agent_config {collaborative_planning, thinking_summaries, visualization}`, background only, ≤60 min), Antigravity coding agent (`antigravity-preview-09-2026`, Linux sandbox 4 vCPU/16 GB, `agent_config {model, max_total_tokens}`, hooks), custom managed agents `POST /v1beta/agents {base_agent, agent_config, system_instruction, tools, base_environment}` (≤1,000/project, no versioning); environments, credentials, triggers (cron), webhooks
**Endpoint:** `POST /v1beta/interactions`
**Params:** `agent`, `agent_config`, `environment`, `background`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW` · `LIVE_VERIFIED`
*Deep Research LIVE_VERIFIED (create → in_progress → cancel); custom agents / triggers / credentials not tested; sandbox compute unbilled in preview* | yes | Four different shapes: agent → session → events (OpenAI, Anthropic), a stateful request loop (xAI), agent-as-model inside Interactions (Gemini).
*Ref:* docs/comparisons/agents-platforms.md · docs/xai/responses.md · docs/xai/grok-build.md · docs/gemini/interactions-api.md |
| **Create an agent session / run**
*4/4* | `POST /v1/agents/sessions {agent\|agent_id, environment (required: none\|openai_hosted\|self_hosted), input, stream, vault_ids}` → 201 `agent.session`
**Endpoint:** `POST /v1/agents/sessions`
**Params:** `agent_id`, `environment`, `input`, `stream`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `POST /v1/sessions {agent, environment_id (required), initial_events[], resources[], vault_ids[], budget, title}` → 200 `session`
**Endpoint:** `POST /v1/sessions`
**Params:** `agent`, `environment_id`, `initial_events`, `budget`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `POST /v1/responses {model, input, tools[], max_turns, store}` → the stored response is the session; continue with `previous_response_id`
**Endpoint:** `POST /v1/responses`
**Params:** `model`, `input`, `tools`, `max_turns`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | `POST /v1beta/interactions {agent\|model, input, background, store, environment, tools, webhook_config}` → `interaction {id, status, steps[]}`; chain with `previous_interaction_id`
**Endpoint:** `POST /v1beta/interactions`
**Params:** `agent`, `input`, `background`, `previous_interaction_id`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*`/v1/interactions` GA UNVERIFIED* | yes | See FAQ Q5 for the body-level comparison.
*Ref:* endpoints.json · docs/faq.md Q5 |
| **Send input / events to a session**
*4/4* | `POST /v1/agents/sessions/{id}/events` with `agent.session.input.message\|tool_result\|cancel` → 202
**Endpoint:** `POST /v1/agents/sessions/{session_id}/events`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `POST /v1/sessions/{id}/events` with `user.message\|interrupt\|tool_confirmation\|custom_tool_result\|tool_result\|define_outcome`, `system.message`
**Endpoint:** `POST /v1/sessions/{session_id}/events`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | a new `POST /v1/responses` with `previous_response_id` (user message or `function_call_output` / `shell_call_output` items)
**Endpoint:** `POST /v1/responses`
**Params:** `previous_response_id`, `input[](function_call_output)`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | a new `POST /v1beta/interactions` with `previous_interaction_id` and `input` (user text or `function_result` steps); `POST …/cancel` to stop
**Endpoint:** `POST /v1beta/interactions`
**Params:** `previous_interaction_id`, `input`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*`interaction.requires_action` for function calls* | yes | Dedicated event endpoints (OpenAI, Anthropic) vs chained requests (xAI, Gemini).
*Ref:* streaming-events.json (client→server) |
| **Session event stream**
*4/4* | SSE via `POST /sessions {stream:true}` or `GET …/events?stream=true`; 31 event types `agent.session.*`, `agent.output.*`
**Endpoint:** `GET /v1/agents/sessions/{session_id}/events`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `GET /v1/sessions/{id}/events/stream` (+ per-thread `/threads/{id}/stream`); 37 event types `agent.*`, `session.*`, `span.*`, `event_start/delta`
**Endpoint:** `GET /v1/sessions/{session_id}/events/stream`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | the ordinary Responses SSE stream (24 recorded events incl. `response.code_interpreter_call.*`, `response.web_search_call.*`, `response.mcp_call.*`) or `wss://api.x.ai/v1/responses`
**Endpoint:** `POST /v1/responses`
**Params:** `stream`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | Interactions SSE (`stream:true`): `interaction.created → interaction.status_update → step.start → step.delta {text\|thought_signature\|arguments_delta\|function_result\|*_call…} → step.stop → interaction.completed → done [DONE]`; resumable with `GET …?stream=true&last_event_id=`; 15 recorded events (5 legacy names RETIRED 2026-06-08)
**Endpoint:** `POST /v1beta/interactions`
**Params:** `stream`, `last_event_id`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | yes |
*Ref:* streaming-events.json |
| **Hosted sandbox**
*3/4* | `environment.type: openai_hosted {packages, setup_commands, network, env, skills, plugins, files, environment_template_id}`; `/workspace`; artifacts from `/workspace/outputs`; ~1 h idle expiry; container rates
**Endpoint:** `POST /v1/agents/sessions`
**Params:** `environment`
**Status:** `DOCUMENTED` · `BETA` | `POST /v1/environments {type: cloud, …}` (Ubuntu 24.04, 8 GB RAM, 10 GB disk, `/workspace`, `/mnt/session/{uploads,outputs}`, `/mnt/memory`); state kept 30 days; $0.08/session-hour
**Endpoint:** `POST /v1/environments`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | no hosted sandbox for agents (the `code_interpreter` tool runs Python without a container object; Grok Build sandboxes locally)
**Endpoint:** `POST /v1/responses`
**Status:** `DOCUMENTED` | `POST /v1beta/environments {sources[] (repository ≤500 MB \| gcs ≤2 GB \| inline ≤1 MB/file), network (unrestricted\|disabled\|allowlist with credentials), from_environment}`; Antigravity sandbox 4 vCPU / 16 GB (Python 3.12, Node 22); files `GET …/files/{path}`, `PUT /upload/…/files/{path}`; idle after 15 min, deleted after 7 days; compute **unbilled during preview**
**Endpoint:** `POST /v1beta/environments`
**Params:** `sources`, `network`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW` · `LIVE_VERIFIED`
*live GET showed `storage.tier: free`, 1 GiB project limit* | yes | Three hosted sandboxes (OpenAI, Anthropic, Gemini); none on xAI.
*Ref:* docs/openai/agents-environments-and-vaults.md · docs/anthropic/managed-agents.md §8 · docs/gemini/interactions-api.md |
| **Self-hosted execution**
*3/4* | `environment.type: self_hosted {workspace_directory}` + `codex exec-server --remote …` executor with environment key `CODEX_API_KEY`; outbound only
**Endpoint:** `POST /v1/agents/sessions`
**Params:** `environment`
**Status:** `DOCUMENTED` · `BETA` | `type: self_hosted` environment = work queue (`/environments/{id}/work`, poll/ack/heartbeat/stop) + worker (`ant beta:worker`, SDK `EnvironmentWorker`) with environment key `sk-ant-oat01-…`
**Endpoint:** `GET /v1/environments/{environment_id}/work/poll`
**Status:** `DOCUMENTED` · `BETA` | client-side by construction: `shell {environment: local}` and `function` tools end the server loop and hand control back; Grok Build CLI runs everything locally
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=shell].environment`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | no self-hosted worker protocol; function calls (`interaction.requires_action`) are the only client-side hook
**Endpoint:** `POST /v1beta/interactions`
**Status:** `DOCUMENTED` | yes | Worker protocols on OpenAI/Anthropic; plain tool round-trips on xAI/Gemini.
*Ref:* endpoints.json (managed-agents) |
| **Credential vaults / secrets**
*3/4* | `/v1/vaults`, `/vaults/{id}/credentials` (rotate/delete); referenced by `vault_ids[]` and `mcp.credential_id`
**Endpoint:** `POST /v1/vaults`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*9 endpoints* | `/v1/vaults`, `/vaults/{id}/credentials` (`mcp_oauth` with refresh, `environment_variable`), `mcp_oauth_validate`; `vault_credential.refresh_failed` webhook
**Endpoint:** `POST /v1/vaults`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*13 endpoints* | pass `authorization` / `headers` on the `mcp` tool per request; no vault
**Endpoint:** `POST /v1/responses`
**Params:** `tools[type=mcp].authorization`
**Status:** `DOCUMENTED` | `/v1beta/credentials` (`bearer_token`, `oauth2`, `environment_variable` with `injection_location`, `trusted_domains`); write-only secrets; referenced from environment network allowlists
**Endpoint:** `POST /v1beta/credentials`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW`
*5 endpoints, not tested* | yes | Three secret stores; xAI inlines credentials.
*Ref:* endpoints.json |
| **Multi-agent / subagents**
*3/4* | `multi_agent {enabled, max_concurrent_subagents (6)}`; `/sessions/{id}/subagents[/{id}/items\|turns]`; items `create_subagent_call`, `wait_for_subagents_call`…; also Responses beta `responses_multi_agent=v1`
**Endpoint:** `GET /v1/agents/sessions/{session_id}/subagents`
**Params:** `multi_agent`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | `multiagent {type: coordinator, agents[] (≤20, incl. self and one advisor)}`; threads (`/sessions/{id}/threads`, ≤25 concurrent); events `agent.thread_message_*`
**Endpoint:** `GET /v1/sessions/{session_id}/threads`
**Params:** `multiagent`
**Status:** `DOCUMENTED` · `BETA` | `grok-4.20-multi-agent-0309` (BETA): the model itself fans out to 4 or 16 agents per `reasoning.effort`; no subagent resources; Grok Build CLI has client-side subagents
**Endpoint:** `POST /v1/responses`
**Params:** `reasoning.effort`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*live output carried `\confidence{80}` markers* | no sub-agents or delegation for custom agents (docs); Deep Research orchestrates internally
**Endpoint:** `POST /v1beta/interactions`
**Status:** `DOCUMENTED` | yes | Orchestration APIs on OpenAI/Anthropic; a multi-agent model on xAI; internal only on Gemini.
*Ref:* docs/comparisons/agents-platforms.md · docs/models/xai-models.md |
| **Session budgets**
*3/4* | no documented budget parameter; turn error code `session_budget_exceeded` exists
**Status:** `DOCUMENTED` · `UNVERIFIED` | `budget {type: limit, max_list_cost {amount (cents), currency: USD}}` enforced between model requests; `session.budget_reached` webhook
**Endpoint:** `POST /v1/sessions`
**Params:** `budget`
**Status:** `DOCUMENTED` · `BETA` | `max_turns` per Responses request (turn budget)
**Endpoint:** `POST /v1/responses`
**Params:** `max_turns`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | Antigravity `agent_config.max_total_tokens` → `status: incomplete`
**Endpoint:** `POST /v1beta/interactions`
**Params:** `agent_config(antigravity).max_total_tokens`
**Status:** `DOCUMENTED` · `PREVIEW` | no | Money (Anthropic), turns (xAI), tokens (Gemini) — not interchangeable.
*Ref:* docs/anthropic/managed-agents.md §13 · docs/xai/responses.md · docs/gemini/interactions-api.md |
| **Outcome grading (define outcome + rubric)**
*1/4* | — not offered | `user.define_outcome` event → grader loop (`max_iterations` ≤20), `span.outcome_evaluation_*` events, `outcome_evaluations[]` on the session
**Endpoint:** `POST /v1/sessions/{session_id}/events`
**Status:** `DOCUMENTED` · `BETA` | — not offered | — not offered | no | Anthropic-only.
*Ref:* docs/anthropic/managed-agents.md §13 |
| **Server-side memory stores**
*1/4* | — not offered | `/v1/memory_stores` (+ memories, versions, redact); mounted at `/mnt/memory`; beta `agent-memory-2026-07-22`; Dreams (`/v1/dreams`, `dreaming-2026-04-21`) reorganise them
**Endpoint:** `POST /v1/memory_stores`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*14 + 5 endpoints* | — not offered | — not offered | no | Anthropic-only.
*Ref:* endpoints.json |
| **Scheduled runs**
*2/4* | — not offered | `/v1/deployments` (cron) → `/v1/deployment_runs`; pause/unpause/run now
**Endpoint:** `POST /v1/deployments`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | — not offered | `/v1beta/triggers {schedule (cron), time_zone, display_name, max_consecutive_failures, execution_timeout_seconds, interaction}`; `PATCH {status: paused\|active}`; `POST\|GET …/executions`
**Endpoint:** `POST /v1beta/triggers`
**Params:** `schedule`, `time_zone`, `interaction`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW`
*7 endpoints, not tested* | yes | Anthropic deployments ≈ Gemini triggers.
*Ref:* endpoints.json |
| **Artifacts / deliverables**
*3/4* | `/sessions/{id}/artifacts[/{id}/content]` from `/workspace/outputs` (≤200 MiB each)
**Endpoint:** `GET /v1/agents/sessions/{session_id}/artifacts`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED` | files written to `/mnt/session/outputs` → `GET /v1/files?scope_id=` + `/content`
**Endpoint:** `GET /v1/files`
**Status:** `DOCUMENTED` · `BETA` | tool outputs (`code_interpreter_call.outputs`, image/video `file_output`) via `include` or the Files API; no artifact resource
**Endpoint:** `POST /v1/responses`
**Params:** `include`
**Status:** `DOCUMENTED` | environment files: `GET /v1beta/environments/{env}/files/{path}?alt=media` (file bytes or tar), legacy `GET /v1beta/files/environment-{env}:download`; Deep Research reports as `model_output` steps (text/image)
**Endpoint:** `GET /v1beta/environments/{environment}/files/{path}`
**Status:** `DOCUMENTED` · `BETA` · `PREVIEW` · `LIVE_VERIFIED`
*partial 200 live* | yes |
*Ref:* docs/openai/agents-environments-and-vaults.md §2 · docs/gemini/interactions-api.md |
| **Embeddable chat UI & workspace agents**
*1/4* | ChatKit (`OpenAI-Beta: chatkit_beta=v1`, `/v1/chatkit/sessions\|threads`) and Workspace Agents (`api.chatgpt.com/v1/workspace_agents/{id}/trigger`)
**Endpoint:** `POST /v1/chatkit/sessions`
**Status:** `DOCUMENTED` · `BETA` · `LIVE_VERIFIED`
*Agent Builder shuts down 2026-11-30* | — not offered | Grok Apps / Grok Bot integrations are consumer products, not an embeddable API (docs/xai/grok-apps-and-integrations.md)
**Status:** `DOCUMENTED` | AI Studio and Firebase AI Logic are builder/SDK surfaces, not an embeddable chat API
**Status:** `DOCUMENTED` | no | OpenAI-only.
*Ref:* docs/openai/chatkit.md · docs/openai/workspace-agents.md |
| **Client-side agent framework / CLI**
*4/4* | Agents SDK (`openai-agents`, `@openai/agents`) — loop over Responses API; handoffs, guardrails, tracing
**Status:** `DOCUMENTED` | Claude Agent SDK / `ant` CLI (`ant beta:sessions`, `ant apply`); SDK `tool_runner` helpers
**Status:** `DOCUMENTED`
*Claude Agent SDK not catalogued in this atlas beyond the migration mapping* | **Grok Build CLI** (`curl -fsSL https://x.ai/cli/install.sh \| bash`; `grok`, `grok -p … --output-format json`, `grok agent stdio` (ACP); `~/.grok/config.toml` `api_backend = chat_completions\|responses\|messages`; MCP servers, hooks, skills, plugins, subagents, Landlock/Seatbelt sandbox; reads CLAUDE.md/AGENTS.md) — BETA; xai-sdk `tool_runner`-style helpers absent
**Status:** `DOCUMENTED` · `BETA` | google-genai SDK automatic function calling (Python callables, `maximum_remote_calls` 10) and `mcpToTool()`; Genkit, Firebase AI Logic, Vercel AI SDK, LangGraph/CrewAI/LlamaIndex integrations; Antigravity is the hosted coding agent
**Status:** `DOCUMENTED` · `BETA` | yes |
*Ref:* docs/openai/agents-sdk.md · docs/anthropic/cli.md · docs/xai/grok-build.md · docs/gemini/sdks.md |
## Model customisation & evaluation
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Fine-tuning / tuning**
*1/4* | `/v1/fine_tuning/jobs` (SFT, DPO, RFT, vision) on gpt-4.1*, gpt-4o*, o4-mini — **DEPRECATED**: no new jobs after 2027-01-06; our key 403 `training_not_available`
**Endpoint:** `POST /v1/fine_tuning/jobs`
**Params:** `model`, `training_file`, `method`
**Status:** `DOCUMENTED` · `DEPRECATED` · `ACCOUNT_RESTRICTED`
*no GPT-5.x/6 fine-tuning* | — not offered | none offered by xAI (capability matrix row 'fine-tuning: none')
**Status:** `DOCUMENTED` | `tunedModels.*` (12 endpoints) still in the v1beta discovery document but **RETIRED** on the Developer API since the Gemini 1.5 Flash-001 deprecation (May 2025): create/list/get → 501 UNIMPLEMENTED; no tunable 3.x/2.x model; use Vertex AI supervised tuning
**Endpoint:** `POST /v1beta/tunedModels`
**Status:** `DOCUMENTED` · `DEPRECATED` · `RETIRED`
*Gemma 4 `tuning: not available`* | no | Only OpenAI still accepts jobs, and only until 2027-01-06.
*Ref:* docs/openai/fine-tuning.md · docs/gemini/tuning.md · docs/models/xai-models.md |
| **Evals**
*1/4* | `/v1/evals`, runs, output items; graders — **DEPRECATED**: read-only 2026-10-31, shutdown 2026-11-30
**Endpoint:** `POST /v1/evals`
**Status:** `DOCUMENTED` · `DEPRECATED` · `LIVE_VERIFIED`
*12 endpoints* | — not offered | — not offered | no evals API on the Developer API (Vertex AI evaluation service)
**Status:** `DOCUMENTED` | no | OpenAI-only (deprecated).
*Ref:* docs/openai/evals.md |
| **Graders (standalone)**
*1/4* | `/v1/fine_tuning/alpha/graders/{run,validate}`
**Endpoint:** `POST /v1/fine_tuning/alpha/graders/run`
**Status:** `DOCUMENTED` · `DEPRECATED` · `BETA` · `LIVE_VERIFIED` | — not offered | — not offered | — not offered | no | OpenAI-only.
*Ref:* docs/openai/graders.md |
| **Stored completions / distillation**
*1/4* | Chat `store:true` + `metadata` → list/retrieve/update/delete stored completions; distillation via SFT
**Endpoint:** `GET /v1/chat/completions`
**Params:** `store`, `metadata`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | — not offered | Chat `store`/`metadata` silently accepted; no stored-completion retrieval endpoints (Responses store instead)
**Endpoint:** `POST /v1/chat/completions`
**Params:** `store`
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED` | — not offered | no | OpenAI-only.
*Ref:* docs/openai/chat-completions.md · parameters.json (xai store) |
| **Reusable prompt templates**
*1/4* | `prompt {id, version, variables}` on Responses — **DEPRECATED**, shutdown 2026-11-30
**Endpoint:** `POST /v1/responses`
**Params:** `prompt`
**Status:** `DOCUMENTED` · `DEPRECATED` | — not offered | — not offered | no prompt registry on the API (AI Studio saves prompts client-side); `cachedContents` reuse a prefix
**Status:** `DOCUMENTED` | no | OpenAI-only (deprecated).
*Ref:* docs/openai/deprecations.md |
## Legacy / retired surfaces
| Feature | OpenAI (how / endpoint / params / status) | Anthropic | xAI | Gemini | Portable | Notes on differences |
|---|---|---|---|---|---|---|
| **Retired agent / thread APIs**
*0/4* | Assistants API **RETIRED 2026-08-26** — 404 empty body; 23 operations mapped to Conversations/Responses/Agents
**Endpoint:** `GET /v1/assistants`
**Status:** `RETIRED` · `DOCUMENTED`
*`OpenAI-Beta: assistants=v2` historical* | — not offered | — not offered | Interactions v1beta legacy schema (`outputs` → `steps`, old SSE names `interaction.start`, `content.*`) removed 2026-06-08; `total_reasoning_tokens` → `total_thought_tokens`
**Endpoint:** `POST /v1beta/interactions`
**Status:** `RETIRED` · `DOCUMENTED` | no |
*Ref:* docs/openai/assistants-retired.md · docs/gemini/interactions-api.md |
| **Retired media models / endpoints**
*0/4* | DALL·E 2/3 RETIRED 2026-05-12; `/v1/images/variations` 404; Sora 2 + Videos API shut down 2026-09-24
**Endpoint:** `POST /v1/images/variations`
**Status:** `RETIRED` · `DEPRECATED` | — not offered | `grok-2-image(-1212)` RETIRED (404); `grok-imagine-image-pro` → `-quality` (DEPRECATED, retires 2026-11-02 → grok-imagine-image-2.0 low); Live Search 410
**Endpoint:** `POST /v1/images/generations`
**Status:** `RETIRED` · `DEPRECATED` | Imagen 3/4 (`:predict`) RETIRED 2026-08-17; Veo 2.0/3.0 RETIRED 2026-06-30; `gemini-2.5-flash-image` DEPRECATED → 2026-10-02; `gemini-omni-flash-preview` → 2026-09-30; half-cascade Live models RETIRED 2025-12-09
**Endpoint:** `POST /v1beta/models/{model}:predict`
**Status:** `RETIRED` · `DEPRECATED` | no |
*Ref:* docs/openai/images.md · docs/xai/deprecations-and-release-notes.md · docs/gemini/deprecations-and-changelog.md |
| **Retired text models**
*0/4* | gpt-3.5-turbo-instruct/babbage/davinci shut down 2026-09-28; o1/o3-mini/o4-mini/gpt-4/gpt-4-turbo 2026-10-23; gpt-5 2025 snapshots, o3 2026-12-11; codex ≤5.2, deep-research, computer-use-preview retired
**Status:** `DEPRECATED` · `RETIRED` | RETIRED on the Claude API: Opus 4.1 (2026-08-05), Sonnet 4 / Opus 4 (2026-06-15), Claude 3.x, 2.x, 1.x, Instant; some still served on Bedrock/Vertex
**Endpoint:** `POST /v1/messages`
**Status:** `RETIRED`
*404 `not_found_error`* | **RETIRED 2026-05-15 with redirects**: `grok-3`, `grok-4-0709`, `grok-4-fast-*`, `grok-4-1-fast-*` → `grok-4.3` (billed at 4.3 rates; `GET /v1/models/grok-3` returns the grok-4.3 object), `grok-code-fast-1` → `grok-build-0.1`; LEGACY/UNVERIFIED: `grok-2-*`, `grok-3-mini`, `grok-beta`, `grok-vision-beta`, `grok-4-latest`
**Endpoint:** `GET /v1/models/{model_id}`
**Status:** `RETIRED` · `LEGACY`
*no consolidated deprecations page; ~60-day notice observed* | Gemini 2.0 Flash/-Lite RETIRED 2026-06-01; 2.5 previews 2025-11/2026-03; `gemini-3-pro-preview` 2026-03-09 (id repointed to 3.1 Pro), `gemini-3.1-flash-lite-preview` 2026-05-25; **Gemini 2.5 Pro/Flash/Flash-Lite 'no longer available to new users'** (404, undocumented); `gemini-3.1-flash-lite` DEPRECATED → 2027-05-07; `gemini-embedding-001` → 2028-05-14; `text-embedding-004` 2026-01-14
**Endpoint:** `POST /v1beta/models/{model}:generateContent`
**Status:** `RETIRED` · `DEPRECATED` · `ACCOUNT_RESTRICTED`
*shut-down ids still appear in `GET /v1beta/models`* | no | Retirement means 404 on Anthropic/Gemini, a dated shutdown on OpenAI, and a **silent redirect** on xAI.
*Ref:* deprecations.json · models.json |
| **Retired / deprecated beta headers, parameters and SDKs**
*4/4* | `OpenAI-Beta: realtime=v1` (legacy `/v1/realtime/sessions` → 404), `assistants=v2`; `prompt_cache_retention`, `truncation: auto`, `user` (→ `safety_identifier`) legacy
**Status:** `LEGACY` · `FAILED_VERIFICATION` | RETIRED: `context-1m-2025-08-07`, `computer-use-2024-10-22`, `max-tokens-3-5-sonnet-2024-07-15`; DEPRECATED: `mcp-client-2025-04-04`; LEGACY (GA, header optional): prompt-caching, message-batches, pdfs, token-counting, files-api, skills, structured-outputs, extended-cache-ttl, code-execution-2025-05-22, interleaved/fine-grained streaming, effort-2025-11-24…
**Status:** `RETIRED` · `DEPRECATED` · `LEGACY` | no headers to retire; DEPRECATED params/features: `max_tokens` (→ `max_completion_tokens`), `logprobs`/`top_logprobs` (ignored ≥4.20), Anthropic-compatible `/v1/messages`, `x_search` per-call billing (→ 2026-09-21), `zdr_status: pii_scrubbing`, Management `teamId` (→ `scope`/`scopeId`); LEGACY: `/v1/completions`, `/v1/complete`, Live Search
**Status:** `DEPRECATED` · `LEGACY` · `RETIRED` | DEPRECATED: `temperature/topP/topK` guidance (2026-07-21), `thinkingBudget` (LEGACY on 3.x), `HARM_CATEGORY_CIVIC_INTEGRITY` (→ `enableEnhancedCivicAnswers`), `googleSearchRetrieval`, **standard API keys** (rejected from Sept 2026), legacy SDKs `google-generativeai` / `@google/generative-ai` (2025-11-30); LEGACY: PaLM methods, `responseSchema`/`responseJsonSchema` (→ `responseFormat`), `mediaChunks`, `speechState`
**Status:** `DEPRECATED` · `LEGACY` | no |
*Ref:* generated/fragments/headers/anthropic-beta-headers.json · deprecations.json (xai, gemini api_features) |
| **Retired model ids keep resolving (redirect aliases)**
*1/4* | retired ids fail with `model_not_found`; `shutdown_date` exposed on `GET /v1/models`
**Endpoint:** `GET /v1/models/{model}`
**Status:** `DOCUMENTED` · `LIVE_VERIFIED` | retired ids → 404 `not_found_error` (some still served on Bedrock/Vertex)
**Endpoint:** `POST /v1/messages`
**Status:** `DOCUMENTED` | retired slugs **redirect** to their replacement and are billed at the replacement's price: `grok-3`, `grok-4-0709`, `grok-4-fast-*`, `grok-4-1-fast-*` → `grok-4.3` (verified: `GET /v1/models/grok-3` returns the grok-4.3 object), `grok-code-fast-1` → `grok-build-0.1`, `grok-imagine-image-pro` → `-quality` → `-2.0 low`; `response.model` reveals the target
**Endpoint:** `GET /v1/models/{model_id}`
**Status:** `DOCUMENTED` · `RETIRED` · `LIVE_VERIFIED`
*6 `retired_redirect` records in models.json* | shut-down ids stay **listed** in `GET /v1beta/models` but generation fails (404); one documented repoint: `gemini-3-pro-preview` → `gemini-3.1-pro-preview` (2026-03-09); `-latest` aliases hot-swap targets
**Endpoint:** `GET /v1beta/models`
**Status:** `DOCUMENTED` · `LIVE_DISCOVERED` | no | xAI-only behaviour (silent redirect); Gemini has one documented id repoint.
*Ref:* deprecations.json (xai) · docs/xai/deprecations-and-release-notes.md |
## Reading the matrix programmatically
```bash
# all features unique to xAI
jq '[.records[] | select(.providers_supporting == ["xai"]) | .feature]' generated/compatibility/cross-provider-feature-matrix.json
# features on all four providers that are marked portable
jq '[.records[] | select(.provider_count == 4 and .portable) | .feature]' generated/compatibility/cross-provider-feature-matrix.json
# every Gemini cell that is ACCOUNT_RESTRICTED (paid-tier feature probed with a free key)
jq '[.records[] | select(.gemini.status|index("ACCOUNT_RESTRICTED")) | {feature, endpoint: .gemini.endpoint}]' generated/compatibility/cross-provider-feature-matrix.json
# every beta-gated Anthropic cell
jq '[.records[] | select(.anthropic.status|index("BETA")) | {feature, endpoint: .anthropic.endpoint}]' generated/compatibility/cross-provider-feature-matrix.json
```
Related pages: [index](index.md) · [models](models.md) · [state management](state-management.md) · [tool execution](tool-execution.md) · [streaming](streaming.md) · [agents platforms](agents-platforms.md) · [pricing](pricing.md) · [caching and reasoning](caching-and-reasoning.md) · [realtime and media](realtime-and-media.md) · [endpoint catalogue](../endpoints/index.md) · [FAQ](../faq.md).