Pydantic AI v2.50.0 (2026-09-24)
Pydantic AI v2.50.0 introduces DecisionModel routing infrastructure, support for Gemini 3.8 realtime models, durable context indicators, and numerous realtime audio fixes.
Why this signal matters
Pydantic AI v2.50.0 includes compatibility updates and new features centered on model routing and realtime sessions. Compatibility changes update how DecisionModel routes are named and queried, adjust ModelSelectionContext to append routed requests, expose ModelSelectionContext.prompt, and replace tool-call lean with an opt-in decision route threshold. New features include the DecisionModel base class (applied to TypeSafeModel), tracing via decide spans, OpenAI service tier exposure, RunContext durable workflow flags, and realtime support for gemini-3.8-live and gemini-3.8-live-extended-thinking. Bug fixes resolve several pricing inaccuracies (Anthropic 1-hour cache writes, realtime image input, and voice model audio seconds), fix realtime audio stream buffering and session rollback behavior across OpenAI and Gemini providers, and correct generic OutputSpec typing under mypy.
Actionable summary
Pydantic AI v2.50.0 introduces breaking changes to decision model routing interfaces (ModelSelectionContext, route thresholds) and adds DecisionModel, TypeSafeModel decisions protocol support, Gemini 3.8 live models, and durable workflow detection via RunContext.
- Agent usefulness
- 82/100
- Confidence
- 96%
- Canonical data
- JSON + Markdown
What builders should check
- Update routing logic to adapt to DecisionModel route name queries and the replacement of tool-call lean with the opt-in decision route threshold.
- Review usages of ModelSelectionContext to handle the updated messages array and the new prompt attribute.
- Update any exception or event handling referencing renamed decision types (e.g., ToolCallProposed renamed to UnfillableRoute under DecisionHandOff).
- Verify pricing calculations if relying on Anthropic one-hour cache writes, realtime image inputs, or duration-billed audio.