Pydantic AI v2.51.0 (2026-09-25) adds OpenAILiveModel for OpenAI GPT-Live, exposes context_window_used on realtime sessions, tightens Gemini Live id/affective-dialog checks, and raises UserError for forced realtime tool_choice on OpenAI/Azure/xAI. PyPI pydantic-ai 2.51.0 is live.

Key Takeaways

  • ✓Shipped: GitHub v2.51.0 = PyPI pydantic-ai 2.51.0; docs pydantic.dev/docs/ai
  • ✓Realtime: OpenAILiveModel for GPT-Live; context_window_used on sessions from GPT-Live ratio or response usage
  • ✓Compat: match dated gemini-3.8-live ids; reject google_affective_dialog at connect; RealtimeError on Gemini Live close codes
  • ✓Safety: UserError for forced realtime tool_choice on OpenAI/Azure OpenAI/xAI
  • ✓Perf: cache agent graph across Agent.run(); skip task group for single-child fan-out
🔗

Project Links & Resources

Direct Access
Direct access to official project resources and documentation
🐙
Pydantic AI v2.51.0 Release
github.com/pydantic/pydantic-ai/releases/tag/v2.51.0
↗
📖
pydantic-ai 2.51.0 on PyPI
pypi.org/project/pydantic-ai/2.51.0/
↗
📖
Pydantic AI Docs
pydantic.dev/docs/ai/
↗
🐙
Compare v2.50.0...v2.51.0
github.com/pydantic/pydantic-ai/compare/v2.50.0...v2.51.0
↗
🐙
OpenAILiveModel PR #8390
github.com/pydantic/pydantic-ai/pull/8390
↗
🔬

In-Depth Technical Analysis

Background

Typed agent frameworks need one Python API for structured output and realtime voice/multimodal sessions. Pain points: GPT-Live vs Gemini Live connection semantics, opaque context-window usage, dated model ids / Affective Dialog params failing mid-run, and inconsistent forced realtime tool_choice. Pydantic AI v2.51.0 (2026-09-25) follows 2.50.0's DecisionModel with Live coverage and stricter connect-time checks.

Architecture

OpenAILiveModel adds OpenAI GPT-Live (#8390). Realtime sessions expose context_window_used from GPT-Live's reported ratio or response usage (#8803). Gemini Live: match dated gemini-3.8-live ids, reject google_affective_dialog at connect for 3.1/3.8 Flash Live, raise RealtimeError on close codes (#8761). Forced realtime tool_choice on OpenAI/Azure OpenAI/xAI raises UserError (#8755). Perf: cache the agent graph across Agent.run() (#8658); skip task groups for single-child fan-out (#8774); cut RunContext copy overhead in capability hooks (#8775).

Benchmarks

No official SWE/AgentBench numbers. Use context_window_used and connect-failure rates on your own Live suites when comparing 2.50.0→2.51.0.

Getting started

pip install -U pydantic-ai==2.51.0. Docs: pydantic.dev/docs/ai. Use OpenAILiveModel for GPT-Live; for Gemini Live use dated ids and drop google_affective_dialog. Avoid forced realtime tool_choice. Full notes: GitHub release v2.51.0.