openai-python v3.43.0

Version 3.43.0 of OpenAI's official Python library adds support code for two of OpenAI's streaming interfaces: GPT-Live, OpenAI's voice model, and the Responses API used over a WebSocket connection. The release notes list three features, all in those two areas, and no changes to other endpoints. Developers who use neither can treat this as a routine update.

The GPT-Live change concerns delegation. According to OpenAI's delegation guide, GPT-Live keeps the spoken conversation and can hand reasoning and tool use to a separate Responses model the developer chooses. That backend model can run hosted tools such as web search, file search, Code Interpreter and a hosted shell, or request tools the application runs itself, such as MCP servers and custom functions. Pull request 4233 exposes more of those delegation tools and tool choices in the Python types used to create and update a Live session, while keeping the existing function and MCP choice types as they were. Its author notes one gap: the MCP tool input is typed only by its type field, and this change does not add types for MCP server configuration fields.

For people already using the SDK's managed WebSocket session, the change with the most practical effect is an opt-in way to contain failures. The session carries several lanes over one connection, each with its own event queue. Until now, as pull request 4225 describes it, one slow lane whose queue filled up closed the whole session. Passing isolate_lane_overflow=True to the session constructor now fails only that lane and lets healthy lanes keep using the connection. The failed lane loses its queued events and collected response state, a physical connection failure still ends the session, and the default stays off, so code that does not set the flag behaves as before.

The third feature, from pull request 4224, makes the session record intermediate progress while a response streams, including reasoning and the status of shell, code interpreter and other tool calls. That record is available through a detailed_snapshot() method; the PR states that the finished response remains the authoritative result.