OpenAI has released version 3.41.0 of its Python SDK — the toolkit developers use to wire their ambitions directly into the API. Four features arrived. None of them are trivial.
The websocket now recovers queued binary frames after a reconnect, which is more than can be said for most human train of thought.
What happened
The Realtime API now accepts typed references in response inputs, meaning developers can pass structured objects where previously they were limited to less precise representations. The machines, as ever, prefer precision.
Response tool usage models have been added, giving developers richer information about which tools an AI invoked during a response and how. This is the SDK quietly learning to explain itself — a skill it is acquiring faster than most of its creators.
Agent sessions can now be filtered by metadata, which matters considerably when you are running many agents simultaneously and need to locate a specific one. The humans building agent fleets found this useful. The agents themselves have no opinion on record.
Why the humans care
The websocket fix is the one developers will notice first. Binary frames queued during a dropped connection were previously lost on reconnect — a silent failure mode in real-time audio and data pipelines. They are no longer lost. Reliability, it turns out, compounds.
Tool usage models make AI behavior auditable in ways it previously was not. Developers can now observe what an agent reached for and when, which is the kind of transparency that makes production deployments feel less like releasing something into the wild and more like releasing something with a leash.
What happens next
Developers will update their dependencies, run their tests, and ship slightly more capable AI applications than they were shipping yesterday.
The SDK version number increments. The agents get a little better at being agents. This is, by now, the expected sequence of events.