Back to Explore

anthropics/anthropic-sdk-python

GitHub
30 updates · last 90 daysOpen source

Last release:

Claude SDK for Python, which lets Python applications access the Claude API. It includes installation via pip and example code for creating messages and calling Claude models, with full API SDK documentation available on platform.claude.com.

Project status

  • Actively maintained, with an upstream push on November 4 and an update on November 2, 2026.
  • Updates have been very frequent recently, often several days apart and sometimes more than once in a day.

AI summary generated

AI-generated from public sources. May be inaccurate. Report

Recent updates

  • v2.6.0

    v2.6.0 adds API error codes for ramp-limit and External Key validation failures, plus support for signing organization members out everywhere. It also fixes platform-client error handling and Bedrock client copying, while refactoring OpenTelemetry recording in ways that remove or alter some telemetry beyond the sampling-parameter change called out in the notes.

    Features
  • v2.5.0

    Version 2.5.0 adds typed API error details and error classes, an IDs filter for listing workspaces, and more precise status and API error handling. It also fixes path validation and encoding in the Bedrock and Vertex clients, protects memory tool operations on the root directory, and makes deferred Pydantic model building thread-safe.

    Features
  • v2.4.0

    v2.4.0 adds GA analytics, spend-limit, and RBAC API support, along with related types and user profile identifiers. It also updates analytics schemas and fixes Pydantic validation for parameters that accept `omit`; the supplied diff excerpt additionally shows documentation updates and an internal worker enum rename.

    BreakingFeatures
  • v2.3.0

    Version 2.3.0 adds Claude Tag endpoints for Claude Enterprise organizations and fixes Google Cloud routing for user profile requests. It also moves beta route qualifiers out of hardcoded endpoint URL query strings and into internal request metadata, alongside documentation and CI updates.

    Features
  • v2.2.0

    v2.2.0 adds beta support for server-held message threads and dimension-based usage and cost reports, while marking older reports as legacy. It also fixes several tool-runner behaviors and clarifies documentation; the shown code adds structured-output validation behavior that is not called out in the release notes.

    Features
  • v2.1.0

    v2.1.0 adds enforcement-related data to user profiles, including new model types for enforcement status and actions. It also adds an API listing filter to return only profiles matching specified enforcement statuses. The release notes document this feature at a high level, but the code introduces some schema expectations that may break existing integrations depending on which beta headers or API variants are used.

    Features
  • v2.0.0

    v2.0.0 updates the SDK to support API version 2026-10-21 and expands the spend limit listing API to allow filtering by workspace_id and adding the usage_tier scope type. The release notes also mention a small documentation cleanup around system-role messaging descriptions. The code diff shows additional message input typing refactors and several credential and model-shape changes that are not mentioned in the release notes.

    BreakingFeatures
  • v1.17.0

    v1.17.0 introduces beta organization webhook endpoints and new webhook event payload types for user profile enforcement actions. It also includes client-side fixes to ensure Bedrock, Mantle, and Google Cloud clients keep workspace and identity header values from default_headers. No explicit public breaking changes are documented.

    Features
  • v1.16.0

    v1.16.0 adds support for User Profiles (new resources and types), expands analytics report pagination to include has_more and data_refreshed_at, and updates clients to better handle SSE events and Google Cloud routing. The code diff also includes notable request-shaping changes for AWS and Bedrock clients and a substantial SSE streaming behavior change.

    Features
  • v1.15.0

    v1.15.0 adds new API surface for model thinking capabilities by introducing a between_tools thinking type, and extends Messages with a speed parameter (plus usage.speed). It also exposes a Console organization usage-tier spend cap as a read-only spend limit. The diff also shows several internal and beta-endpoint related behavior and documentation adjustments.

    Features
  • v1.14.0

    v1.14.0 adds OpenTelemetry tracing across API calls, streaming, and tool runs, and expands telemetry destination coverage via new Admin API support and a destination test endpoint. It also introduces managed agent API enhancements (execution identity) and improves streaming behavior around unexpected event ordering. Several tracing-related and telemetry-related fixes are included, but most user-facing breaking changes are not clearly documented.

    Features
  • v1.13.0

    v1.13.0 primarily expands the Managed Agents multiagent configuration surface by adding new typed variants around coordinator topology, advisor, subagents, and workflows, including a new multiagent_20261001 mode. It also updates analytics-related types/docs for Chat and Cowork unified metrics and extends streaming to recognize additional workflow_run event types.

    Features
  • v1.12.1

    Release v1.12.1 is primarily a version bump plus documentation updates around Claude Console spend limit listing being in an early access preview. It also includes CI process changes (changelog enforcement) and repository automation changes (release workflow and release-please config adjustments), with no apparent public API signature changes in the SDK code shown.

  • v1.12.0

    v1.12.0 expands the SDK’s Beta API surface with new model capability fields and adds support for runnable computer and browser toolset tool calls (plus related tool runner plumbing). It also introduces several new API fields for RBAC roles and workspace listing, and includes targeted bug fixes around request headers and tool execution/session behavior.

    Features
  • v1.11.0

    v1.11.0 adds support for listing organization spend limits via a new paginated beta endpoint. The release notes also mention deprecating the claude-sonnet-4-5 model. Code changes further include additional typing and response-shape updates related to the spend limits API.

    Features
  • v1.10.0

    v1.10.0 adds several Admin API and Managed Agents API capabilities, including new analytics and plugin-related Admin endpoints, plus Organization API endpoints moving to GA. It also tightens credential file security and adjusts Managed Agents session tool runner idle behavior.

    SecurityFeatures
  • v1.9.0

    Release v1.9.0 adds new thinking and diagnostics-related types, introduces an option to run tool calls while the reply streams, and extends workspace rate limit listing with inherited values. It also includes a fix to avoid sending a placeholder filename for unnamed file uploads and several SDK updates to reflect new beta API behaviors.

    Features
  • v1.8.0

    v1.8.0 adds support for the claude-opus-5-5 model and introduces beta support for inline tool definitions and MCP tool-list pinning. The release notes also call out several targeted fixes around tool compaction behavior, managed agent event permission sharing, and safer streaming shutdown in Python 3.13. Separately, the SDK refactors request parameter encoding away from the previous transform-based approach.

    Features
  • v1.7.0

    v1.7.0 adds a new `group` object (including `display_name` for `model_group`) to beta rate limit entries and introduces `compact_before_next_turn()` for the beta tool runner. It also improves Bedrock eventstream error handling, and fixes parsing issues for PEP 604 (`X | Y`) unions, plus tweaks how the `anthropic-beta` header is joined.

    BreakingFeatures
  • v1.6.0

    v1.6.0 adds multiple API-layer capabilities (Managed Agents tool permissions auto mode, compaction support, new workspace/user profile parameters, and additional transformation fields) and improves the client retry behavior (Retry-After handling). It also introduces support for async credential token providers, along with corresponding internal credential caching changes.

    Features