OpenAI’s Python SDK has a new release with five additions for developers building agent workflows. The changes include running local tools during streamed session creation and clearer usage information for agent turns, making the SDK more useful in the places where an agent actually does work.
OpenAI Watch analysis
What happened
Version 3.25.0 of the OpenAI Python SDK was published on 6 October. Its release notes list five features for agents, the API and tool helpers, alongside a public type export.
Our top picks
- Run local tools during streamed session creation
Agents can use local tools while a streamed session is being created. - Observe local tool failures
The SDK adds a way for agents to observe when local tool calls fail. - Represent agent turns in the API
New agent turn items provide API support for representing those turns. - Group usage by source
Usage information can now be grouped by source, giving developers a more useful breakdown. - Search tools in typed function helpers
Typed function helpers now support tool search with deferred loading enabled. The v3.25.0 release notes also make typetoresponseformatparam publicly available.
Why it matters
These are practical building blocks for developers using the SDK to connect agents with local tools and track what happens across a turn. Tool failures are easier to surface, streamed sessions can run local tools, and usage can be separated by source rather than left as one undifferentiated total.
Our read
This is a useful, focused SDK release: less grand unveiling, more plumbing that can make agent applications easier to build and inspect. If you work with the OpenAI Python SDK, the release notes are worth checking against your tool and usage-tracking setup.
What to watch
- Whether the new tool-search support fits existing typed function workflows.
- How developers use the added turn and usage details in agent applications.
- What the next SDK release changes for local tool handling.
Discussion spark: For agent developers, which matters more in practice: being able to run local tools during streamed sessions, or getting clearer failure and usage details?
Sources and evidence
- v3.25.0 (6 October 2026, 17:56 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.