OpenAI Python SDK v3.15.0 Adds Managed WebSocket Sessions and Prompt-Cache Prewarming

OpenAI shipped version 3.15.0 of its official Python client library on September 18, 2026, adding five new API features centered on agent sessions, streaming infrastructure, and caching performance.

What's New

The release introduces managed Responses WebSocket sessions, allowing developers to maintain persistent, stateful connections to the Responses API rather than issuing discrete HTTP calls. This is paired with new compaction progress events, which expose real-time status updates during context compaction operations — useful for long-running agent sessions that need to manage token budgets dynamically.

OpenAI also added agent session model settings, giving developers finer control over model configuration within persistent agent sessions, and prompt-cache prewarming, a feature that lets applications proactively warm the prompt cache before inference requests arrive, aimed at reducing latency on cache-dependent workloads.

Rounding out the feature list, the SDK now exposes audio-mini model choices, expanding the set of smaller, presumably lower-cost audio models available through the API. OpenAI has not disclosed pricing or specific model names for these audio-mini options in the release notes.

Bug Fixes and Chores

The update fixes a bug where chat stream moderation results were not preserved across streaming responses, which could have caused moderation flags to be dropped in certain streaming configurations. Two chore items clarify documentation: usage guidance for incoming SIP call IDs (relevant to OpenAI's realtime voice/telephony features) and updated image request examples. A separate documentation fix removes nonexistent type imports from the realtime module.

What this means

This is a client-library changelog, not a new model release — no new weights, parameter counts, or benchmark scores are involved. The substantive signal here is infrastructure-focused: OpenAI is investing in making agent sessions more stateful and controllable (managed WebSocket sessions, agent session model settings) and in reducing latency for high-volume or repeated-prompt workloads (prompt-cache prewarming, compaction progress events).

For developers building agentic applications on top of the Responses API, the WebSocket session support and compaction progress events are the most actionable additions — they suggest OpenAI is optimizing for longer-running, multi-turn agent workloads where connection overhead and context management have been pain points. The audio-mini model choices hint at an expansion of OpenAI's audio model lineup, though details on pricing and capability remain undisclosed pending a separate model announcement. None of these changes require action from developers using standard chat completions, but teams building persistent agents or voice applications should review the updated SDK documentation.