Source: raw/anthropic-watch-anthropic-sdk-python-tag-v1-{1..9}-0.md and raw/anthropic-watch-anthropic-sdk-typescript-tag-{sdk-v0-121-0,sdk-v0-128-0,sdk-v0-129-0,bedrock-sdk-v0-33-2,bedrock-sdk-v0-34-0,vertex-sdk-v0-20-0,foundry-sdk-v0-5-0}.md (Stainless-generated GitHub releases, 2026-08-26 → 2026-09-28), plus the anthropics/skills commits 53048666, 34040c9c, 33375500 and 8a1541c4 that document the same API surface in the claude-api skill, and cookbook commit 6b671ef6 (Admin API notebook).
Project: anthropics/anthropic-sdk-python, anthropics/anthropic-sdk-typescript · URL: https://github.com/anthropics/anthropic-sdk-python/releases · Version: Python v1.1.0 → v1.9.0; TypeScript sdk-v0.121.0 → sdk-v0.129.0
The Python SDK shipped nine minor releases in the five weeks after its breaking 1.0.0 (2026-08-20, covered in the May–July SDK article), and the TypeScript SDK stayed on 0.x. Two model IDs landed: claude-opus-5-5 (Python v1.8.0 / TS sdk-v0.128.0, 2026-09-22) and claude-sonnet-5-5 (Python v1.9.0 / TS sdk-v0.129.0 and the Bedrock, Vertex and Foundry packages, 2026-09-28). The other headlines are cache diagnostics going GA, auto-mode tool permissions for Managed Agents, Organization API endpoints, a compaction beta, and a tool runner that can start tool calls while the reply is still streaming.
Key Takeaways
- Bump to Python v1.8.0 / TS sdk-v0.128.0 for Opus 5.5, and v1.9.0 / sdk-v0.129.0 for Sonnet 5.5. For Sonnet 5.5 on the cloud providers, the TypeScript packages are
bedrock-sdk-v0.34.0,vertex-sdk-v0.20.0andfoundry-sdk-v0.5.0. See Claude Opus 5.5 and Claude Sonnet 5.5. - Cache diagnostics are GA (Python v1.9.0 / TS sdk-v0.129.0): a
diagnosticsfield onMessageandMessageCreateParams. The beta dates to Python v0.102.0 / TS sdk-v0.96.0 in May. Theclaude-apiskill’s cost-optimize flow now points cache misses at it (skills8a1541c4). See Prompt Caching for Agencies. - Three breaking or tightening changes in minor releases — check these before upgrading:
webhooks.unwrap()requires headers (Python v1.2.0).- Beta user profiles removed
relationshipin favour ofaccess_type(Python v1.3.0). - Credentials files readable by group or others are refused (Python v1.5.0).
- Managed Agents got an
autopermission policy (Python v1.5.0, repeated in v1.6.0). Theclaude-apiskill describes it as a third option besidealways_allow/always_ask: the action runs, is denied as high-risk with an error tool result while the session keeps running, or pauses for approval when indeterminate — and warns it is “not a human checkpoint” (skills34040c9c). See Managed Agents. - The tool runner grew up across the period. It keeps going on
pause_turn(v1.1.0), gainedcompact_before_next_turn()(v1.7.0), madeadd_tools()/addTools()take effect straight away (v1.8.0 / sdk-v0.128.0), and can optionally run tool calls while the reply streams (v1.9.0 / sdk-v0.129.0). - Files and Skills APIs moved to GA shapes. Python v1.2.0’s beta files/skills namespaces use GA shapes and drop dated beta header pins; the
claude-apiskill says both APIs “are out of beta” and its examples callclient.files.*/client.skills.*with no beta header (skills53048666). - TypeScript deprecated the
betasparam on GA models and completions methods, and removed client-side compaction control from its tools (sdk-v0.129.0, both filed as chores).
Python v1.1.0 → v1.9.0
| Version | Date | Highlights |
|---|---|---|
| v1.1.0 | 2026-08-26 | ”updates” thinking display mode (beta); Organization API endpoints; tool runner keeps going on pause_turn; legacy response API removed (filed as an internal chore) |
| v1.2.0 | 2026-08-27 | Beta files/skills namespaces use GA shapes and drop dated beta header pins; webhooks.unwrap() requires headers; AWS/Bedrock sign raw request bytes so binary uploads work; agent toolset and memory tool preserve exact file bytes |
| v1.3.0 | 2026-09-01 | Beta user profiles: external_user_onboarded_at added, relationship removed in favour of access_type; organization compliance settings; batch results on GA raw/streaming response wrappers; an unexplained “api: manual updates” entry |
| v1.4.0 | 2026-09-04 | Claude Tag category and user breakdowns in usage reports; workspace ID on more endpoints; clear error when an httpx object is passed instead of httpx2 |
| v1.5.0 | 2026-09-10 | Auto mode tool permissions for Managed Agents; content_too_large web_fetch error; public GitHub repos mount without authorization_token in Managed Agents sessions; Message.to_param(); tool objects accepted directly in messages.create/parse/stream/count_tokens; credentials files readable by group/others refused; idempotency_key kept as a deprecated no-op |
| v1.6.0 | 2026-09-15 | compaction parameter and signed compaction blocks (beta); thinking_mismatch_allowed in input_transformations (beta); url_sources on web fetch; async credential token providers; honours Retry-After above 60 seconds; async client retries connection errors |
| v1.7.0 | 2026-09-18 | Rate limits gain group with display_name, group_type deprecated; tool runner compact_before_next_turn(); Bedrock raises an API error for eventstream exception frames; multiple anthropic-beta values joined with a comma and no space; requires pydantic 1.10+ |
| v1.8.0 | 2026-09-22 | claude-opus-5-5; inline tool definitions and MCP tool-list pinning (beta); one shared evaluated_permission enum across Managed Agents events; Dreams API reference docs |
| v1.9.0 | 2026-09-28 | claude-sonnet-5-5; between_tools thinking type (degrades to disabled on fallback hops); cache diagnostics GA; tool runner can run tool calls while the reply streams; workspace rate limits include_inherited and source; typed event-type filter for Managed Agents events; parse and tool runner stop sending the beta header |
The Organization API endpoints in v1.1.0 are the client.beta.organization surface that the cookbook’s new Admin API notebook (2026-09-18) exercises: invites and roles, workspaces and members, an API key audit, service accounts and rate limits (cookbook 6b671ef6). See Claude Cookbooks.
TypeScript
sdk-v0.121.0(2026-08-26) — the same feature set as Python v1.1.0 (updates thinking display beta, Organization API endpoints, tool runner onpause_turn), plus Standard Schema support for structured outputs and tools. Examples switched to adaptive thinking.sdk-v0.128.0(2026-09-22) —claude-opus-5-5, inline tool definitions and MCP tool-list pinning (beta);addTools()takes effect straight away; the tool runner’s compaction request leaves out reply-only params.sdk-v0.129.0(2026-09-28) —claude-sonnet-5-5,between_toolsthinking, aClientToolUniontype for client-executed tools, cache diagnostics GA, optionally start tool calls while the reply streams, and a partial-JSON streaming fix that shows every complete array item and holds back unfinished numbers. Chores: deprecate thebetasparam on GA models and completions methods, and remove client-side compaction control from tools.bedrock-sdk-v0.33.2(2026-08-26) — never sends ambient first-party credentials to the Mantle endpoint.bedrock-sdk-v0.34.0,vertex-sdk-v0.20.0,foundry-sdk-v0.5.0(2026-09-28) — each addsclaude-sonnet-5-5.
The claude-api skill as release notes
The anthropics/skills commits document API behaviour the SDK changelogs only name:
- Fable 5.1 / Mythos 5.1 (skills
53048666, 2026-09-01):claude-fable-5-1andclaude-mythos-5-1become the default Fable-tier IDs; per-message effort (mid-conversation-output-config beta); turn-scoped mid-conversation system messages withclear_at;thinking.display "updates"for progress between tool calls (apparently the “updates” display mode the SDKs added in v1.1.0 / sdk-v0.121.0^[inferred]); thinkingblock_bindingcontrols; a 0.025x cache-read rate on Fable 5.1 with amax_tokens: 0keep-alive. See Claude Fable 5.1 + Mythos 5.1. - Managed Agents
autopolicy (skills34040c9c, 2026-09-10): theevaluated_permissionandevaluationfields onagent.tool_use/agent.mcp_tool_use; client code should gate onevaluated_permission === 'ask'rather than the configured policy;ant beta:sessions connectterminal viewer; the deny field isdeny_message, notmessage. - Refusal billing (skills
33375500, 2026-09-24): a refusal before any output counts against rate limits; a mid-stream refusal bills input tokens plus the output already streamed. - Opus 5.5 / Sonnet 5.5 (skills
8a1541c4, 2026-09-29): Opus 5.5 is the skill’s default model and Sonnet 5.5 the current Sonnet; streaming client tools default toeager_input_streaming; model access failures are a 404 and beta gating a 400, not a 403;ant applyis recommended for version-controlled Managed Agents resources.
Try It
- Pin
anthropic>=1.9.0(Python) or@anthropic-ai/sdk@0.129.0(TypeScript) to get both new model IDs and cache diagnostics GA. - Before upgrading Python past v1.2.0, grep for
webhooks.unwrap(and pass the request headers. - Before upgrading past v1.5.0, check that your credentials file is not readable by group or others (
chmod 600). - Read
diagnosticson responses to find out why a prompt-cache read missed, instead of inferring it fromusage. - If you run Managed Agents unattended, evaluate the
autopolicy — and treat it as a classifier, not a reviewer.
Open Questions
- Did a
claude-fable-5-1model constant land in any SDK? No SDK release in this batch names it. Python v1.3.0 (2026-09-01, the Fable 5.1 launch day) carries an unexplained “api: manual updates” entry. - TypeScript sdk-v0.122.0 through sdk-v0.127.0 are not in these sources (sdk-v0.128.0 diffs against v0.127.0), nor are
bedrock-sdkv0.33.3–v0.33.8 (v0.34.0 diffs against v0.33.8). - What are “inline tool definitions” and “MCP tool-list pinning”? Both are one-line beta entries in v1.8.0 / sdk-v0.128.0 with no documentation link.
- What does
between_toolsthinking do? The changelogs add the type and say it degrades to disabled on fallback hops; its behaviour is not described. - Will TypeScript follow Python to 1.0? Still no announcement.
Related
- Anthropic SDK Releases — May–July 2026 — the preceding article, including Python v1.0.0.
- Claude Opus 5.5 —
claude-opus-5-5. - Claude Sonnet 5.5 —
claude-sonnet-5-5. - Managed Agents — the
autopermission policy. - What’s New — Week 39 and What’s New — Week 40 — the Claude Code side of the two launches.
- Claude Cookbooks — the Admin API organization-management notebook.
- Prompt Caching for Agencies — what cache diagnostics measure for.