# xAI (Grok) models catalogue

Status: LIVE_VERIFIED for the 12 ids returned by GET https://api.x.ai/v1/models on 2026-09-19 (all 7 language models also exercised with one minimal completion each; image/video models listed only, no media generated); DOCUMENTED for voice models (not in the catalogue endpoint), ACCOUNT_RESTRICTED for grok-embedding-small, RETIRED (+ redirect verified live) for the May-15-2026 batch, LEGACY/UNVERIFIED for grok-2/grok-3-mini/grok-beta era ids.
Sources: https://docs.x.ai/developers/models · /developers/models/ (per-model pages, fetched live 2026-09-19 — absent from llms-full.txt) · /developers/pricing · /developers/rate-limits · /developers/release-notes · /developers/grok-4-6 · /developers/model-capabilities/text/{reasoning,multi-agent} · /developers/migration/{may-15-retirement,imagine-image-quality-nov-2} · /developers/rest-api-reference/inference/models · sources/xai/{models-api-raw,language-models-raw,image-generation-models-raw}.json · tmp-live/xai-models/catalogue-probe*.json.
Last verified: 2026-09-18 (UTC 2026-09-19) · Machine-readable twin: generated/fragments/models/xai-models.json (built by scripts/gen_xai_fragments.py); prices in generated/fragments/pricing/xai-pricing.json.

# 1. Live catalogue (GET /v1/models, 12 ids)

max_output: xAI publishes no output-token limit ("No text output limit" on the grok-4.6 page); none is returned by the catalogue.

# Aliases (live aliases[])

Alias convention (docs): <model> = latest stable, <model>-latest = latest version, <model>-<date> = pinned. In practice the dated 4.20 ids are the canonical ids and the bare names are aliases.

# 2. Documented ids outside the catalogue

# 3. Live vs docs discrepancies (2026-09-19)

• Price units: the REST reference says token prices are "USD cents per 100 million tokens" and image_price is "USD cents"; live values are consistent with ticks of 1e-10 USD for both (12500 → $1.25/1M tokens, 200000000 → $0.02/image). price_per_image is documented as "1/100,000,000ths of a USD cent" = the same unit. cost_in_usd_ticks (usage) uses the same 1e10 ticks/USD.
• grok-4.5: model page lists xhigh as supported; reasoning page says only low/medium/high and xhigh is treated as high.
• grok-4.20-0309-reasoning, grok-4.20-0309-non-reasoning, grok-build-0.1: model pages show "Reasoning: Yes/No" but reasoning_effort → 400 invalid-argument "Model … does not support parameter reasoningEffort". Only grok-4.6, grok-4.5, grok-4.3 (and multi-agent via reasoning.effort) accept it.
• grok-4.20-multi-agent-0309: POST /v1/chat/completions → 400 "Multi Agent requests are not allowed on chat completions" (bare string body); works on POST /v1/responses. The response's reasoning.effort defaulted to "medium"; output text carried a trailing \confidence{80} / \confidence{90} marker; usage contained undocumented context_details{input_tokens,output_tokens}.
• grok-imagine-video-1.5: live input_modalities = text, image, audio; model page says text, image.
• grok-4.5 created 2026-06-29 (live) vs July 2026 release note; grok-build-0.1 created 2026-04-16 vs May 2026 note.
• Regional catalogues: GET https://us.api.x.ai/v1/models → only grok-4.6 with prices ×1.1 (22000 / 5500 / 66000 ticks, long-context 44000 / 11000 / 132000). GET https://eu-west-1.api.x.ai/v1/models → 200, only grok-4.3 at global prices — this host is not documented (model pages mention the eu-west-1 cluster only).
• search_price is 0 for every language model (Live Search superseded by tools).
• GET /v1/api-key returned empty strings for create_time / modify_time (docs: Unix timestamp; example: RFC 3339).
• Documentation stubs: the gRPC reference pages and per-model pages are rendered client-side — llms-full.txt contains only their titles. Per-model pages were fetched live as .md; gRPC methods were taken from the xai-sdk 1.19 proto stubs (see docs/xai/grpc-api.md).

# 4. Model × Capability matrix (text models)

# 5. Model × Endpoint matrix

# 6. Model × Tool matrix (type strings)

Responses API (tools[].type): web_search, x_search, code_execution (alias code_interpreter), image_generation, attachment_search, collections_search (alias file_search), view_image / view_x_video (implicit, on search results), mcp (remote MCP), function (client). gRPC / xai-sdk: web_search(), x_search(), code_execution(), collections_search(), mcp() — code_interpreter and file_search aliases unsupported. Documentation lists tool support explicitly only for grok-4.6 (function calling, web search, X search, code execution); all text models accept function tools. Details: docs/tools/xai/index.md.

# 7. Pricing summary (USD per 1M tokens; full detail in docs/xai/pricing.md)

Tiered rule: when the prompt reaches 200k tokens all tokens of the request are billed at the long-context rate. Image input tokens cost the same as text. Reasoning tokens are output tokens. Images: 1.0 $0.02, quality $0.05, 2.0 $0.04 default (low 1k/1.5k/2k = 0.04/0.05/0.06; medium = 0.06/0.07/0.08). Video: 1.0 $0.05/s, 1.5 $0.08/s.

# 8. Rate limits (Tier 0, documented → observed)

Full tier tables: docs/xai/rate-limits.md.

# 9. Observed usage shapes (minimal "Reply with OK.", max_tokens 8)

Chat usage = {prompt_tokens, completion_tokens, total_tokens, prompt_tokens_details{text_tokens, audio_tokens, image_tokens, cached_tokens}, completion_tokens_details{reasoning_tokens, audio_tokens, accepted_prediction_tokens, rejected_prediction_tokens}, num_sources_used, cost_in_usd_ticks} plus top-level system_fingerprint, service_tier: "default", message.reasoning_content, message.refusal. Responses usage adds num_server_side_tools_used, input_tokens_details.cached_tokens, output_tokens_details.reasoning_tokens, context_details. Note the ~190–640 prompt tokens for a 4-token prompt (system template overhead; FAQ acknowledges API counts exceed the tokenizer).