OpenAI’s official Python library for building applications that use the OpenAI API. Useful for developers who want to authenticate and make API requests from Python.
Project status
- Actively maintained: The upstream GitHub repository shows a very recent push today (2026-09-10), and the library has multiple new version updates within the last few days, indicating ongoing maintenance rather than dormancy.
- Apparent update cadence: Recent updates include v3.10.0 on 2026-09-09, v3.11.0 on 2026-09-09, and v3.9.0 on 2026-09-08, followed by continued updates in late August 2026, suggesting a fast, ongoing cadence (at least weekly and often more frequent).
AI summary generated
Recent updates
v3.11.0
v3.11.0 introduces expiration controls for service account keys. The main code change adds new request/response fields related to API key expiration for admin organization service accounts and project API keys, along with type and resource method updates.
Featuresv3.10.0
v3.10.0 primarily extends the client API surface for GPT Image 2.5 models and adds service-account API key expiration support. It updates request/response types for image generation (new model names and quality/size options) and adds expires_in_seconds/expires_at to the service accounts API key creation flow.
Featuresv3.9.0
v3.9.0 adds prompt cache diagnostics support, including new response fields and request options to compare cache reuse behavior. It also expands web search tool call status handling to include an incomplete state, and hardens retry-after parsing to avoid pathological server delays.
Featuresv3.8.0
v3.8.0 adds support for the new gpt-6-astra model and introduces additional beta API capabilities (notably around response steering over WebSocket). Release notes also mention documentation work for a canonical SDK security model, but the code contains other functional additions.
SecurityFeaturesv3.7.0
Release v3.7.0 updates the admin organization usage costs API documentation and adds a parameter for filtering costs by specific line items. It also improves embedding parsing performance by avoiding repeated numpy capability checks. Several response typing/schema fields were changed as part of the OpenAPI/spec update, including usage and incomplete-response reason handling.
Featuresv3.6.0
v3.6.0 adds compute_units to the token usage models returned by Responses and Chat Completions. It also includes a hardening pass for X.509 workload identity authentication, plus related schema updates and dependency/tooling bumps.
Featuresv3.5.0
v3.5.0 adds an API change that relaxes the requirement for function tool call output objects to include a call identifier. Specifically, the `call_id` field is no longer required and can be omitted or null in both response schemas and beta schemas.
BreakingFeaturesv3.4.0
Release v3.4.0 adds support for an obfuscation field in ChatCompletion streaming chunks and introduces project residency configuration support for the admin projects API. The release also includes multiple bug fixes around Realtime call encoding, SSE decoding, and redirect and auth boundary handling for WebSockets and Azure.
Featuresv3.3.1
v3.3.1 mainly updates dependencies (including security fixes) and migrates the project build and contributor tooling from Rye to uv. The only runtime library code change shown is in platform detection logic in src/openai/_base_client.py, alongside version bumps and various repo/tooling workflow updates.
BreakingSecurityv3.3.0
v3.3.0 introduces support for named data-residency endpoints (global, us, eu, ae) for OpenAI clients by allowing regional routing selection. The release also updates optional networking dependencies to enforce patched security floors and removes/pins some build tooling dependencies.
BreakingFeaturesv3.2.0
v3.2.0 adds Bedrock Runtime endpoint support and introduces new beta streaming event types for shell calls, along with related service and image types. The release notes only document those two areas, but the code diff shows several additional behavioral and type-level changes across the client and models.
Featuresv3.1.0
v3.1.0 adds new Realtime/WebSocket capabilities (including WebSocket stream IDs), introduces a new workload identity audit-log event, and expands service tier support with an Ultrafast tier. It also adds structured MCP and websocket error types and separates websocket error events.
Featuresv3.0.0
Release v3.0.0 migrates the SDK implementation to HTTPX2 as the default HTTP layer. The code diff shows HTTPX2 types are threaded through client construction, request preparation, response handling, and exposed error/response objects.
Breakingv2.54.0
This release (v2.54.0) adds new Responses model identifiers and clarifies requirements for audio upload metadata. The diff also shows repository-wide code generation header attribution moving from Stainless to Castiron, plus several CI and release workflow changes.
Featuresv2.53.0
v2.53.0 adds support for the gpt-5.5 model and extends Responses tool-call typing to include the tool output’s name and namespace. The release notes also mention a CI fix to avoid NumPy source builds and duplicate HTTPX coverage.
Featuresv2.52.1
Release v2.52.1 updates the package version metadata and adjusts CI configuration. The only functional change called out in the release notes is a CI workflow pin for setup-uv v5.
v2.52.0
v2.52.0 adds a new content provenance checks API (with both sync and async client support) and introduces better retry handling for server-directed Retry-After delays, extending the honored window to up to two minutes. It also updates documentation with new API-key mTLS HTTP client recipes.
Featuresv2.51.0
This release (v2.51.0) extends the existing `service_tier` parameter to accept a new value, `fast`, and updates request/response typing across chat completions and responses APIs. Release notes mention adding the fast tier and wiring it into helper methods, but the code diff contains additional behavioral documentation not present in the notes.
Featuresv2.50.0
v2.50.0 updates transcription-related API surface and models. The release notes call out a transcription keyword overload fix, plus general transcription model updates.
Featuresv2.49.0
v2.49.0 raises the supported Python floor to 3.10 (from 3.9) and adds automation to continuously review and validate Python version support. The release also introduces new CI workflows and scripts to enforce wheel metadata and version-policy consistency, and updates internal tooling and dependency lockfiles accordingly.
BreakingFeatures