Pydantic AI Updates: September 19, 2026

1. Pydantic AI 2.46.0 Let TypeSafeModel Fill Tool Arguments and Union Outputs

Pydantic AI shipped 2.46.0 a day after 2.45.0, extending the TypeSafeModel it had just added for TypeSafe’s Jev. The new release lets TypeSafeModel fill a tool’s arguments when Jev can express them, and choose among a union of output types rather than being limited to a single shape. A typesafe_boolean_threshold setting controls how confident a yes/no has to be before it resolves to True, which makes the calibrated-probability output of that model class usable as a gate rather than a hint. Source

2. The Release Also Added a Model Profile Flag for Judge-Only Models

Pydantic AI added supports_text_output to ModelProfile, which lets LLMJudge and GEval run against a model that does not emit free text at all. That is the practical consequence of models like Jev entering the stack: evaluation code that assumed a text completion now has to branch on capability. The release also added RealtimeSession.wait_for_playback() so realtime examples can finish speaking a reply before the session closes. Source