OpenAI Python SDK v3.20.0 Released: Agents Credentials, Session Options, and WebSocket Enhancements

OpenAI's official Python SDK has landed a v3.20.0 update. This release centers on expanding Agents capabilities, real-time WebSocket optimizations for the Responses API, and a focused round of connection stability fixes for Realtime/Live scenarios. For developers building AI applications with this SDK, this is an iteration aimed squarely at engineering reliability and real-time interaction quality.

New Features: Expanded Agents and Responses Capabilities

This release introduces several noteworthy additions at the API level. The SDK adds credential and session configuration options for Agents (#3967), giving developers more flexible control over authentication and session state when building Agent-based workflows. This enables finer-grained management for complex scenarios involving multi-turn conversations and multi-tool calls.

On the Responses API side, Cyber access programs support has been added (#3956), further broadening the interface's feature coverage.

For real-time interactions, Responses gains two WebSocket-related improvements: opt-in support for receiving incremental WebSocket text and tool call snapshots (#3973), and more detailed WebSocket accumulator snapshot retention (#3981). Together, these improvements serve a single goal — making intermediate states in streaming output more complete and observable, helping developers achieve finer-grained real-time rendering and debugging.

The WebSocket accumulator snapshot is a key mechanism in real-time streaming scenarios. During streaming, the server splits responses into incremental delta chunks and pushes them progressively. The accumulator is responsible for assembling these chunks into a complete intermediate state. Retaining more detailed snapshots means developers can read a full view of all content received so far at any point in time — not just the latest delta. This is especially valuable when you need to render, parse, or apply conditional logic mid-stream — for example, detecting tool call intent before the full arguments have arrived, or triggering sentence-by-sentence playback in text-to-speech scenarios.

Bug Fixes: Real-Time Connection Stability Takes Center Stage

If the new features are icing on the cake, the bug fixes in this release are what truly address developer pain points. The fix list is heavily concentrated on connection reliability for both Live and Realtime WebSocket scenarios.

At the client level, the SDK now retries on unmapped TLS transport errors (#3982), reducing the chance of requests failing outright due to underlying network jitter.

Several tricky Live scenario issues were resolved: avoiding hangs near fractional transcription group cutoff times (#3970); stripping query parameters from WebSocket endpoint paths (#3972) to prevent path concatenation errors; and preserving the caller queue to prevent non-deterministic WebSocket replays (#3980).

The Realtime scenario also received hardening, including preserving base URL query parameters during WebSocket upgrades (#3971), and retaining configured queues without replaying previously attempted send requests (#3978). The common theme across these fixes is ensuring deterministic message queue state during connection upgrades, reconnections, and replays — preventing duplicate sends or data loss.

A WebSocket Upgrade is the process by which a client initiates a protocol switch via a standard HTTP request, upgrading the connection from HTTP/HTTPS to a persistent, bidirectional WebSocket channel. This handshake phase is relatively fragile: if query parameters on the upgrade request are mishandled during path construction, or if the queue of unsent messages isn't properly restored after a failed upgrade and reconnect, data loss or duplicate sends can occur. These fixes specifically target such edge cases — ensuring query parameters in the base URL are fully preserved during the upgrade, and that message queue state is deterministic after reconnection. This fundamentally eliminates sporadic message ordering and replay anomalies that are notoriously difficult to reproduce and diagnose in high-concurrency or unstable network production environments.

Documentation: More Systematic Error Response Coverage

The Chores section of this release is almost entirely dedicated to cleaning up API error response documentation. The SDK now provides clearer and more complete error response documentation across multiple interfaces, covering batch errors (#3961), file and upload errors (#3960), fine-tuning and model errors (#3964), Responses not-found errors (#3959), stored chat completion errors (#3963), and a general clarification of API error responses (#3965).

While these changes don't affect code behavior, their practical value to developers shouldn't be underestimated — clear error documentation means faster issue diagnosis and lower debugging overhead.

What This Means for Developers

Taken together, v3.20.0 is a classic "polish" release. There are no earth-shattering new features, but it delivers solid work in two directions: on one hand, Agents credential/session options and new Responses features expand the toolkit for building sophisticated applications; on the other, the large volume of WebSocket and real-time connection fixes directly improve stability for voice and streaming interaction scenarios.

For teams already running this SDK in production — especially developers building real-time voice or conversational applications on top of the Realtime or Live APIs — an upgrade evaluation is recommended. Queue state and replay-related fixes have a way of resolving intermittent, hard-to-reproduce production issues.

Upgrading is straightforward — a pip update to 3.20.0 will get you all the improvements.