SPB Git forge
spboucher.ai ↗

spb/doc-api

Public
2commits 1branches 0releases
15.7 MBsize
maindefault branch
13 days agolast push
Python 88.3% TypeScript 7.6% Shell 4.1%
15.8 KB
View source Raw Blame History

# Anthropic Claude Platform release notes — API-surface digest (2025 → 2026-09-18)

Status: DOCUMENTED (digest of the official release notes; newest first as in the source, grouped by month).
Sources: https://platform.claude.com/docs/en/release-notes/overview (same content as release-notes/api; retrieved 2026-09-18). Model launch dates cross-checked against GET /v1/models created_at (see docs/models/anthropic-models.md §2 for the small offsets).
Last verified: 2026-09-18.

Focus: changes to the API surface (endpoints, parameters, headers, tools, models, limits, pricing). Console-only and Claude-apps items are omitted or one-lined. Bold = model launch/retirement.

# 2026

# September

  • 09-18 Compliance API local sessions also return Claude in Chrome transcripts (product_surface: claude_in_chrome, beta).
  • 09-14 Compact on demand: top-level compaction param → signed compaction block (beta compact-2026-09-04).
  • 09-10 Managed Agents permission policy auto (server evaluates each tool call; evaluation field on agent.tool_use events); ant CLI 1.32 ant beta:sessions connect.
  • 09-03 ant apply (resources as code, claude-lock.json); per-message effort beta also on Google Cloud (Fable 5.1, Mythos 5.1, Opus 5).
  • 09-01 — Claude Fable 5.1 (claude-fable-5-1) and Claude Mythos 5.1 (claude-mythos-5-1, Glasswing): 1M context default, 128k output, always-on adaptive thinking, $10/$50, cache reads $0.25 (0.025x). tool_choice any/tool → 400 on these models. Thinking blocks bound to producing model (older models can't read them); prefix check on replay (400 for accounts created ≥ 2026-08-31); beta thinking-binding-controls-2026-08-01 adds input_transformations + thinking.block_binding.prefix_mismatch_behavior. Per-message effort beta (mid-conversation-output-config-2026-07-01). Turn-scoped system messages beta (mid-conversation-system-clear-at-2026-08-21, clear_at: next_user_message). thinking.display: "updates" beta (thinking-display-updates-2026-08-18). Text watermark + C2PA credentials on code-execution media via Files API. 30-day retention required. Admin/Analytics/Compliance docs now show anthropic-version — send it everywhere.

# August

  • 08-27 SDKs (Py 1.2.0, TS 0.122.0, Go 1.68.0, Java 2.59.0, Ruby 1.67.0, C# 12.44.0): client.beta.files / client.beta.skills stop sending files-api-2025-04-14 / skills-2025-10-02; BetaSkill → BetaContainerSkill; beta.skills.delete deletes all versions. Console: personal keys and service-account keys (workspace keys legacy).
  • 08-26 Compliance API session endpoints GA for Cowork/Claude Code; Claude Science and Microsoft 365 surfaces (beta). Admin API in CLI + all SDKs under client.beta.organization (usage/cost reports and CE endpoints stay curl-only).
  • 08-20 Python SDK v1.0: httpx → httpx2; Python ≥ 3.10; removes Text Completions API, temperature/top_p/top_k on Messages, tool runner compaction_control; async .with_raw_response needs await response.parse(); AnthropicBedrock errors without region. Computer/browser toolsets on Google Cloud (Fable 5, Mythos 5, Opus 5, Sonnet 5, Opus 4.8).
  • 08-19 Computer use GA as computer_toolset_20260801 (no header, batch actions, zoom default, configs); browser use tool browser_toolset_20260801 launched; both for Fable 5, Mythos 5, Opus 5, Sonnet 5, Opus 4.8. Files API GA (expires_in_seconds/expires_at, page/next_page, ids[]; header → old shape). Agent Skills / /v1/skills GA. Admin API CE user management GA (ce-user-management-2026-07-13 optional). Managed Agents web_search/web_fetch domain allow/block lists in agent_toolset_20260401 configs; memory stores in self-hosted sandboxes.
  • 08-18 Workbench → playground.
  • 08-11 Compliance API local sessions (GET /v1/compliance/apps/sessions/local…, beta). New response header anthropic-workspace-id.
  • 08-10 Sonnet 5 introductory $2/$10 made permanent.
  • 08-07 Managed Agents: session budgets (budget_reached stop), advisor entry in multiagent roster, model.inference_geo, skills from GitHub .claude/skills.
  • 08-05 Inference hooks (CE, beta). Claude Opus 4.1 retired (claude-opus-4-1-20250805).
  • 08-03 Compliance API remote Cowork sessions (beta). 08-01 Dreams supports Opus 5.

# July

  • 07-24 — Claude Opus 5 (claude-opus-5): 1M context (default and max), 128k output, thinking on by default, $5/$25. Breaking: thinking: disabled only at effort ≤ high (400 at xhigh/max). Full effort ladder incl. max. Mid-conversation tool changes beta (mid-conversation-tool-changes-2026-07-01) on Fable 5, Mythos 5, Opus 4.8, Opus 5. fallbacks "default" mode (server-side-fallback-2026-07-01). Fast mode removed for Opus 4.7 (error).
  • 07-22 Managed Agents: model.effort; environment/memory-store webhooks; initial_events on POST /v1/sessions; optional version on agent update (409 on mismatch); thread-level event_deltas[].
  • 07-17 Legacy Workbench and /v1/experimental/{generate,improve,templatize}_prompt retire 2026-08-17.
  • 07-15 Mid-conversation system messages available on Fable 5, Mythos 5, Opus 4.8 (API, Bedrock, Vertex; no header).
  • 07-14 Admin API CE user management beta (ce-user-management-2026-07-13 for groups/roles).
  • 07-10 Dreams supports Fable 5 and Sonnet 5. 07-08 API-key expiration (expires_at in Admin API).
  • 07-02 agent-memory-2026-07-22 header (memory list semantics; replaces managed-agents-2026-04-01 on /v1/memory_stores; both → 400); SDKs send it by default.
  • 07-01 Access to Fable 5 / Mythos 5 restored.

# June

  • 06-30 — Claude Sonnet 5 (claude-sonnet-5): $2/$10 intro, 1M context, 128k output, same tools as Sonnet 4.6 minus Priority Tier; adaptive thinking on by default; manual enabled thinking → 400; non-default sampling params → 400; new tokenizer (~+30% tokens). Managed Agents: event_deltas[] on session stream, prev_page on GET /v1/sessions, agent_with_overrides, vault injection_location, agent/deployment webhooks.
  • 06-29 Fast mode removed for Opus 4.6 (silently standard; usage.speed).
  • 06-26 Rate limits raised; tiers consolidated to Start/Build/Scale; Sonnet/Haiku = Opus limits.
  • 06-25 Fast mode on Opus 4.7 deprecated (removal 07-24). 06-22 MCP tunnels API moved to /v1/tunnels (mcp-tunnels-2026-06-22, workspace:manage_tunnels).
  • 06-18 SDK support for code_execution_20260120 (REPL persistence; minimum for programmatic tool calling).
  • 06-15 Sonnet 4 and Opus 4 retired. 06-11 code_execution_20260521 (90-s per-cell limit disclosed); web_search_20260318 / web_fetch_20260318 add response_inclusion.
  • 06-10 GET /v1/environments/{id}/work on Claude Platform on AWS.
  • 06-09 — Claude Fable 5 (claude-fable-5) and Claude Mythos 5 (claude-mythos-5): 1M default, 128k, always-on adaptive thinking (no disabled, no budgets, no prefill → 400); new tokenizer; safety classifiers → stop_reason: "refusal" (unbilled if no output); opt-in fallbacks (beta, not on Batches); stop_details.category adds reasoning_extraction; thinking.display defaults to omitted; 30-day retention. Managed Agents scheduled deployments; vault env-var credentials; Compliance Activity Feed on Claude Platform on AWS; session_thread_id in webhooks; Swift package for Apple Foundation Models (beta). Mythos Preview deprecated (same day).
  • 06-05 Opus 4.1 deprecation announced (retire 08-05). 06-02 Advisor tool max_tokens; refusals with no output are no longer billed.

# May

  • 05-29 Managed Agents webhooks/multiagent/self-hosted sandboxes on Claude Platform on AWS.
  • 05-28 — Claude Opus 4.8 (claude-opus-4-8): 1M default (API, Bedrock, Vertex, Foundry), 128k, same tools as 4.7. Mid-conversation system messages launched (role: "system" in messages, no header). stop_details (category: cyber/bio/null + explanation) documented. Effort default high; min cacheable 1,024; high-res images; task budgets, advisor, computer use support 4.8; fast mode on 4.8 (research preview); sampling params → 400. Fast mode on Opus 4.6 deprecated.
  • 05-27 usage.output_tokens_details.thinking_tokens (final message_delta when streaming).
  • 05-19 MCP tunnels research preview; Managed Agents self-hosted sandboxes; update MCP/tool config of active sessions; >100K-char tool outputs spilled to a sandbox file.
  • 05-18 Web search returns richer SEC filing data. 05-13 Cache diagnostics public beta (diagnostics.previous_message_id → cache_miss_reason; cache-diagnosis-2026-04-07).
  • 05-12 Fast mode on Opus 4.7. 05-11 — Claude Platform on AWS launched (Claude API on Anthropic-managed infra via AWS endpoints, IAM auth, AWS billing/CCU).
  • 05-06 Managed Agents multiagent orchestration + Outcomes public beta; mcp_oauth credential refresh; webhooks; session/event filters. Dreams research preview (dreaming-2026-04-21).
  • 05-04 Workload Identity Federation (OIDC tokens → POST /v1/oauth/token).

# April

  • 04-30 context-1m-2025-08-07 retired for Sonnet 4.5 and Sonnet 4 (no effect; >200k errors).
  • 04-29 Claude API skill (open-source Agent Skill). 04-24 Rate Limits API. 04-23 Managed Agents memory public beta. 04-20 Claude Haiku 3 retired.
  • 04-16 — Claude Opus 4.7 (claude-opus-4-7): $5/$25, new tokenizer, breaking changes vs 4.6 (see migration guide); Claude in Amazon Bedrock open to all (/anthropic/v1/messages, 27 regions); task budgets beta (task-budgets-2026-03-13); high-res images (2576 px long edge, ~3x image tokens); xhigh effort.
  • 04-14 Sonnet 4 / Opus 4 deprecation announced (retire 06-15). 04-09 Advisor tool public beta (advisor-tool-2026-03-01).
  • 04-08 Claude Managed Agents public beta (managed-agents-2026-04-01); ant CLI launched.
  • 04-07 Claude Mythos Preview (gated, Glasswing); Messages API on Amazon Bedrock research preview (/anthropic/v1/messages, us-east-1).

# March

  • 03-30 Batches max_tokens cap 300k for Opus 4.6 / Sonnet 4.6 (output-300k-2026-03-24); 1M beta retirement for Sonnet 4.5/4 announced for 04-30.
  • 03-18 Models API capability fields: max_input_tokens, max_tokens, capabilities on GET /v1/models and /v1/models/{id}.
  • 03-16 thinking.display ("omitted" keeps signature). 03-13 1M context GA (no header) for Opus 4.6 / Sonnet 4.6 at standard pricing; dedicated 1M rate limits removed; media limit 100 → 600 images/PDF pages per request at 1M.

# February

  • 02-19 Automatic caching (top-level cache_control). Sonnet 3.7 and Haiku 3.5 retired (Haiku 3.5 stays on Bedrock/Vertex). Haiku 3 deprecation announced (retire 04-20).
  • 02-17 — Claude Sonnet 4.6 (claude-sonnet-4-6): extended thinking, 1M context (beta). Code execution free with web search/web fetch. Web search and programmatic tool calling header-free; dynamic filtering for web search/fetch; code execution, web fetch, tool search, tool use examples, memory tool header-free.
  • 02-07 Fast mode research preview on Opus 4.6 (speed param, up to 2.5x).
  • 02-05 — Claude Opus 4.6 (claude-opus-4-6): adaptive thinking recommended (enabled deprecated), no prefill; effort GA (no header) and on Opus 4.6; compaction API beta; data residency (inference_geo, US ×1.1 for models after 2026-02-01); 1M beta on Opus 4.6 (long-context pricing >200k at the time); fine-grained tool streaming header-free.

# January

  • 01-29 Structured outputs GA on Claude API (Sonnet 4.5, Opus 4.5, Haiku 4.5): output_format → output_config.format; still beta on Bedrock/Foundry.
  • 01-12 console.anthropic.com → platform.claude.com redirect. 01-05 Claude Opus 3 retired.

# 2025

# Q4

  • 12-19 Haiku 3.5 deprecation announced. 12-04 Structured outputs on Haiku 4.5.
  • 11-24 — Claude Opus 4.5 (claude-opus-4-5-20251101); programmatic tool calling public beta; tool search tool public beta; effort parameter public beta (Opus 4.5); client-side compaction in Py/TS SDK tool_runner.
  • 11-21 search_result blocks on Bedrock header-free. 11-19 Docs moved to platform.claude.com/docs. 11-18 Claude in Microsoft Foundry launched.
  • 11-14 Structured outputs public beta (structured-outputs-2025-11-13; Sonnet 4.5, Opus 4.1).
  • 10-28 Sonnet 3.7 deprecation announced; Sonnet 3.5 models retired; context editing adds clear_thinking_20251015.
  • 10-16 Agent Skills (skills-2025-10-02; /v1/skills; requires code execution). 10-15 — Claude Haiku 4.5 (claude-haiku-4-5-20251001).

# Q3

  • 09-29 — Claude Sonnet 4.5 (claude-sonnet-4-5-20250929); global endpoint pricing on Bedrock/Vertex; new stop reason model_context_window_exceeded; memory tool beta; context editing beta (clear_tool_uses).
  • 09-17 Tool helpers / tool runner beta (Py, TS). 09-16 Rebrand to Claude Console / Claude Docs / Claude Help Center (APIs unchanged).
  • 09-10 Web fetch tool beta; Claude Code Analytics API. 09-08 C# SDK beta. 09-05 Rate-limit charts in Console. 09-03 Citable documents in client tool results. 09-02 Code execution v2 (Bash + files) public beta. 08-27 PHP SDK beta.
  • 08-26 Higher 1M rate limits for Sonnet 4; 1M on Vertex. 08-19 request_id in error bodies. 08-18 Usage & Cost API; Admin API GET /v1/organizations/me.
  • 08-13 Sonnet 3.5 deprecation announced (retire 10-28); 1h cache TTL GA. 08-12 1M context beta for Sonnet 4 (API + Bedrock). 08-11 Acceleration limits now 429 (were 529). 08-08 search_result blocks GA (API, Vertex). 08-05 — Claude Opus 4.1 (claude-opus-4-1-20250805) (cannot set both temperature and top_p).
  • 07-28 text_editor_20250728 (max_characters). 07-24 / 07-17 Rate limits raised for Opus 4 / Sonnet 4. 07-21 Claude 2.0, 2.1, Sonnet 3 retired. 07-03 search_result blocks beta (search-results-2025-06-09).

# Q2

  • 06-30 Opus 3 deprecation announced. 06-11 Fine-grained tool streaming beta (fine-grained-tool-streaming-2025-05-14).
  • 05-22 — Claude Opus 4 and Sonnet 4: summarized thinking with encrypted signature; interleaved thinking beta (interleaved-thinking-2025-05-14); Files API beta; code execution tool beta; MCP connector beta; top_p default 0.999 → 0.99 (0.95–1 allowed with thinking); Go SDK stable. 05-21 Ruby SDK stable. 05-07 Web search tool. 05-01 cache_control must be on the parent tool_result / document.source block.
  • 04-09 Ruby SDK beta.

# Q1

  • 03-31 Java SDK stable; Go SDK beta.
  • 02-27 URL sources for images and PDFs; tool_choice: none; tools optional when replaying tool_use/tool_result; OpenAI-compatible endpoint.
  • 02-24 — Claude Sonnet 3.7 (extended thinking); vision on Haiku 3.5; token-efficient tool use; bash_20250124, text_editor_20250124 (no header), computer_20250124 (computer-use-2025-01-24).
  • 02-10 anthropic-organization-id response header. 01-31 Java SDK beta.
  • 01-23 Citations; plain-text and custom-content documents. 01-21 Claude 2 / 2.1 / Sonnet 3 deprecation announced. 01-15 Caching reads longest cached prefix automatically; prefill with tools. 01-10 Caching optimized for Batches.

# 2024 (context)

12-19 Batches delete endpoint · 12-17 Models API, Batches, token counting, prompt caching, PDF support GA (no headers); Java/Go SDK alpha · 11-21 Admin API · 11-20 TPM → ITPM/OTPM · 11-04 Haiku 3.5 (text-only) · 11-01 token counting, PDF support · 10-22 computer use tools + Sonnet 3.5 (Oct) · 10-08 Batches beta; relaxed turn ordering · 10-03 disable_parallel_tool_use · 09-10 Workspaces · 08-22 CORS/browser SDK · 08-19 8k output default Sonnet 3.5 · 08-14 prompt caching beta · 06-20 Sonnet 3.5 · 05-30 tool use GA.