# Anthropic models catalogue

Status: DOCUMENTED for every id below; LIVE_VERIFIED for the 11 ids returned by GET /v1/models on 2026-09-18 (each also probed with GET /v1/models/{id}; 5 of them with a minimal POST /v1/messages), ACCOUNT_RESTRICTED for the invite-only Mythos ids (404 with our key), RETIRED (+ 404 verified) for retired ids.
Sources: models/overview, models//overview, about-claude/pricing, about-claude/model-deprecations, about-claude/models/model-ids-and-versions, api/models/list, build-with-claude/{thinking,effort,fast-mode,context-windows,prompt-caching,batch-processing,structured-outputs,vision}, agents-and-tools/tool-use/ (all under https://platform.claude.com/docs/en/), plus sources/anthropic/models-api-raw.json and tmp-live/anthropic-models/summary.json.
Last verified: 2026-09-18 · Machine-readable twin: generated/fragments/models/anthropic-models.json (built by scripts/gen_anthropic_models_fragment.py).

# 1. Current lineup (documented, active or deprecated)

Aliases (pre-4.6 only): claude-opus-4-5 → claude-opus-4-5-20251101 (GET /v1/models → 200); claude-sonnet-4-5 → claude-sonnet-4-5-20250929 (GET /v1/models → 200); claude-haiku-4-5 → claude-haiku-4-5-20251001 (GET /v1/models → 200). Dateless 4.6+ ids are pinned snapshots, not aliases.

# 2. Live vs docs discrepancies (2026-09-18)

• claude-fable-5-1: live created_at 2026-08-28 != documented release 2026-09-01
• claude-fable-5: live created_at 2026-06-07 != documented release 2026-06-09
• claude-opus-4-7: live created_at 2026-04-14 != documented release 2026-04-16
• claude-opus-4-6: live created_at 2026-02-04 != documented release 2026-02-05
• claude-sonnet-5: live created_at 2026-06-29 != documented release 2026-06-30
• claude-sonnet-4-5-20250929: live max_input_tokens 1000000 != documented context window 200000
• claude-haiku-4-5-20251001: live capabilities.code_execution.supported=false while code-execution-tool docs list claude-haiku-4-5-20251001 as supported (without PTC/REPL persistence)
• claude-mythos-5-1, claude-mythos-5, claude-mythos-preview: documented (invite-only) but GET /v1/models/{id} → 404 not_found_error with our key and absent from GET /v1/models (does not mean non-existent).
• GET /v1/models returns undocumented-but-referenced fields max_input_tokens, max_tokens, capabilities{batch,citations,code_execution,context_management{clear_tool_uses_20250919,clear_thinking_20251015,compact_20260112},effort{low..max},image_input,pdf_input,structured_outputs,thinking{types{enabled,adaptive}}} (added 2026-03-18). The reference example shows max_input_tokens: 0; live values are real numbers.
• Messages usage (live) contains cache_creation{ephemeral_5m_input_tokens,ephemeral_1h_input_tokens}, output_tokens_details.thinking_tokens (absent on Sonnet 4.6), service_tier: "standard", inference_geo: "global"; top level container: null, stop_details: null. No speed field unless fast mode is requested.

# 3. Model × Capability matrix

code_execution row = live capabilities.code_execution.supported; docs list Haiku 4.5 as supported (without programmatic tool calling / REPL persistence).
sampling params: temperature/top_p/top_k non-default → 400 on Claude 4.7+ and Mythos Preview; incompatible with thinking on older models.

# 4. Model × Tool-version matrix (exact type strings)

Legacy tool versions: computer_20250124 (beta computer-use-2025-01-24) for Sonnet 4.5, Haiku 4.5, Opus 4.1, Sonnet 4, Opus 4; computer_20241022/bash_20241022/text_editor_20241022 (beta computer-use-2024-10-22) only for Sonnet 3.5 (Oct 2024, retired); text_editor_20250124 for Sonnet 3.7 (retired); text_editor_20250429 Claude 4.x (superseded by text_editor_20250728); code_execution_20250522 Python-only (legacy); mcp-client-2025-04-04 deprecated. Undated aliases tool_search_tool_regex / tool_search_tool_bm25 resolve to the latest dated version. Advisor tool: executor→advisor pairs are constrained (advisor must be at least as capable; see advisor-tool page).

# 5. Model × Endpoint matrix

# 6. Model × Cloud matrix

Bedrock: Opus 4.7+ and Haiku 4.5 use the Messages-API endpoint ids (anthropic.claude-…, no -v1:0); Opus 4.6 is the last -v1 id (anthropic.claude-opus-4-6-v1); dated models use anthropic.claude-{name}-{YYYYMMDD}-v1:0 via InvokeModel. Vertex: dated models use @YYYYMMDD. Foundry and Claude Platform on AWS use the Claude API id (alias form where one exists). Partner clouds set their own retirement dates and +10% regional-endpoint pricing.

# 7. Retired models (deprecations page + release notes)

# 8. Pricing snapshot (USD per MTok; full detail in docs/anthropic/pricing.md)

# 9. Notes

• Model IDs: 4.6+ ids are dateless pinned snapshots (claude-{name}-{major}[-{minor}]); pre-4.6 ids carry -YYYYMMDD and have convenience aliases. Weights never change under an id; serving infrastructure may.
• Tokenizer: Claude 4.7+ and Mythos Preview use a new tokenizer (~30% more tokens for the same text); 1M tokens ≈ 555k words on it vs ≈ 750k words before.
• Thinking: adaptive is the only mode on Fable/Mythos 5.x (cannot be disabled; no budgets; no prefill); Opus 5 / Sonnet 5 have thinking on by default and reject type: enabled; Opus 4.6 / Sonnet 4.6 accept enabled but it is deprecated; 4.5 models are manual-budget only.
• Fast mode: only Opus 5 and Opus 4.8 (speed: "fast" + fast-mode-2026-02-01); Opus 4.7 → error since 2026-07-24; Opus 4.6 → silently standard since 2026-06-29.
• Rate-limit buckets: Fable 5.x shared; Mythos 5.x shared; Opus 5 own; Opus 4.8/4.7/4.6/4.5 shared; Sonnet 5 own; Sonnet 4.6/4.5 shared; Haiku 4.5 own.
• Data retention: Fable 5/5.1 and Mythos 5/5.1 require 30-day retention (no ZDR unless authorized).
• Watermarking: text from Fable 5.1 / Mythos 5.1 carries Anthropic's text watermark; code-execution media files carry C2PA credentials via the Files API.