OpenAI Python SDK v3.12.0 Adds Live API Support, Fixes Async Streaming Bugs

OpenAI Python SDK v3.12.0 Adds Live API Support

OpenAI released version 3.12.0 of its official Python client library on September 10, 2026. The update adds support for a new Live API endpoint and resolves three bugs affecting streaming and type handling. This is a software development kit (SDK) update, not a new model release.

What changed

New feature:

• Live API support — the SDK now includes client bindings for OpenAI's Live API, allowing developers to call the endpoint directly through the Python library. OpenAI's release notes do not specify the underlying model, latency characteristics, or pricing associated with the Live API in this changelog.

Bug fixes:

• Added aclose() method to AsyncStream objects, giving developers a standard way to explicitly clean up async streaming connections instead of relying on garbage collection or context managers (#2854).
• Fixed handling of bare dict and list type annotations that lack type arguments (e.g., dict instead of dict[str, int]), which previously caused errors during type resolution (#3760).
• Fixed a bug where finalized output from a response could be discarded if the completion event arrived as null, ensuring partial or completed results are preserved correctly (#3345).

Confirmed vs. unconfirmed

The existence of the Live API endpoint and these three bug fixes are confirmed by the GitHub release commits and linked pull requests. OpenAI has not published accompanying documentation in this release describing the Live API's capabilities, supported models, context window, or pricing — those details are not yet disclosed in the source material reviewed.

What this means

This is a client library update, not a model launch. The addition of Live API bindings signals that OpenAI is preparing or has already shipped a Live API product, but developers won't get functional details — model backing, latency, pricing per 1M tokens — until OpenAI publishes API documentation separately. The aclose() fix matters for production systems: without it, long-running async streaming applications risked leaking connections during shutdown, a subtle bug that could degrade service reliability under load. The dict/list annotation fix and null-completion handling are narrower correctness fixes that primarily affect developers using strict type checking or handling edge cases in streaming responses. None of these changes require action unless teams are actively hitting the specific bugs described or want early access to Live API bindings.