# Anthropic — provider research

**Last documentation audit: 2026-09-08** — sources: Claude Platform docs (Messages API, Models API, extended thinking,
structured outputs, tool use, prompt caching, pricing, errors), the bundled `claude-api` reference (models table cached
2026-06-24), SDK type definitions of `@anthropic-ai/sdk` 0.124.0, and live probes with the owner key (`scripts/provider-matrix.ts`
and `GET /v1/models`).

## Endpoint, auth, SDK
- Base URL `https://api.anthropic.com`; headers `x-api-key`, `anthropic-version: 2023-06-01`. Invalid key → **401** `authentication_error`.
- SDK `@anthropic-ai/sdk` 0.124.0 (`client.messages.create({stream:true})` returns an async iterable of raw events; `client.models.list()` auto-paginates; `client.messages.countTokens`).
- Timeouts: SDK default 10 min; PolyLLM uses 10 min for chats, 20 s for validation/listing; SDK retries 1 (PolyLLM never retries mid-stream).

## Models API (live capability metadata — verified)
`GET /v1/models` returns `id, display_name, created_at, max_input_tokens, max_tokens, capabilities` with
`thinking.types.{adaptive,enabled}`, `effort.{low,medium,high,xhigh,max}`, `image_input`, `pdf_input`, `structured_outputs`,
`code_execution`, `citations`, `batch`, `context_management`. 11 models for this key: `claude-fable-5-1`, `claude-opus-5`,
`claude-sonnet-5`, `claude-fable-5`, `claude-opus-4-8`, `claude-opus-4-7`, `claude-sonnet-4-6`, `claude-opus-4-6`,
`claude-opus-4-5-20251101`, `claude-haiku-4-5-20251001`, `claude-sonnet-4-5-20250929`. Aliases (`claude-haiku-4-5`) resolve to dated snapshots.
All current models: 1M context (Haiku 4.5: 200k), 128k max output (Haiku 4.5: 64k).

## Verified rules encoded in the adapter
- **Adaptive-only generations** (`thinking.types.enabled.supported === false`: Fable 5/5.1, Opus 5/4.8/4.7, Sonnet 5) reject
  `temperature`/`top_p`/`top_k` (400) and `budget_tokens`; the adapter exposes none of them. Effort via `output_config.effort`.
  Sending `thinking:{type:"adaptive"}` to Haiku 4.5 → 400 `adaptive thinking is not supported on this model` (probed) → Haiku uses `{type:"enabled", budget_tokens}` (≥1024, < max_tokens).
- Fable 5.x: thinking always on, `{type:"disabled"}` → 400; forced `tool_choice` `any`/`tool` → 400 on Fable 5.1 (adapter downgrades to `auto`).
- Adaptive thinking decides when to think: trivial prompts produce **no thinking block and 0 thinking tokens** even at effort `high` (probed on Sonnet 5); a harder prompt produced 3.5k thinking tokens with `display:"summarized"`.
- Streaming events: `message_start` (input usage incl. cache), `content_block_start/delta/stop` with `text_delta`, `thinking_delta`,
  `signature_delta`, `input_json_delta`, `citations_delta`; `message_delta` carries `stop_reason` (`end_turn|max_tokens|stop_sequence|tool_use|pause_turn|refusal|model_context_window_exceeded`),
  `usage.output_tokens` and `usage.output_tokens_details.thinking_tokens`; `stop_details` on refusal.
- Tools: `tools[{name, description, input_schema, strict}]`; tool results as `tool_result` blocks in a **user** turn; thinking blocks replayed with signature only on the same model.
- Server tools: `web_search_20260209` (Opus 5/4.8/4.7/4.6, Sonnet 5/4.6; not combinable with `code_execution_20260521`), `web_search_20250305` for older models.
- Structured output: `output_config.format = { type: "json_schema", schema }` (Opus 4.7+, Sonnet 5, Fable — `structured_outputs.supported`).
- Files: PDF as `document` (base64), text files as `document` with `text/plain` source; images base64 `image` blocks.
- Errors: `{type:"error", error:{type, message}}`: `authentication_error`, `permission_error`, `not_found_error`, `rate_limit_error`, `overloaded_error` (529, retryable), `billing_error`, `request_too_large`, `invalid_request_error`.
- Pricing (USD/M, first-party): Fable 5.1 10/50 (cache read 0.25), Opus 5 5/25, Sonnet 5 2/10, Sonnet 4.6 3/15, Haiku 4.5 1/5 — see `catalog.ts`; cache reads 10 % of input.
- Prompt caching: `cache_control` (not used by PolyLLM v0.1; usage reports `cache_read_input_tokens`, folded into cached tokens).

## Probe results (2026-09-08, `pnpm providers:matrix`)
Auth ✅ · Model list ✅ (11) · Text ✅ · Streaming ✅ · System ✅ · Vision ✅ (64×64 PNG → "Red") · Tools ✅ (streamed `input_json_delta`, args rebuilt) ·
Structured output ✅ · Reasoning ✅ (Sonnet 5, effort high, 3 558 thinking tokens, summary streamed) · Token usage ✅ · Invalid key ✅ (401) ·
Error normalization ✅ (unknown model → `not_found_error` → MODEL_NOT_FOUND) · Long response ✅ (407 words).