Pydantic AI 2.64 adds model fallback controls and artifact scanning

Pydantic AI 2.64 introduces a ModelFallback capability with per-attempt request preparation and retry controls. A sticky option keeps a conversation on the fallback model that answered, while first_event_timeout allows a stream that never starts to trigger fallback.

Keep provider state with its origin

Stored response, item, conversation and container identifiers are sent back only to the endpoint that created them, through Model.accepts_server_state. This matters when an application changes providers or falls back: an identifier from one server is not automatically meaningful to another. The release also changes exhausted transport error handling so eligible HTTP failures can reach fallback logic.

Scan artifacts and choose a workspace

ArtifactScanner adds security scanning for memory and skills, and BoxdSandbox runs harness capabilities in a boxd machine. Scanning is a new control, not a guarantee that an artifact is safe. Sandbox provisioning and access remain application responsibilities.

Other additions include on-demand capability loading, realtime speech word timestamps for supported services, and plain-function evaluators. Durable online evaluations are dispatched once per run rather than on each replay.

Review compatibility notes

json-repair becomes optional and is needed only for specific capabilities. Coder’s strict tool schemas become opt-in. Nested repository instruction files are surfaced more fully, which can change the instructions an agent receives. Existing applications should check these defaults, state routing and fallback behavior against their own tests; the release does not include provider inference credits.

See the pydantic-ai hub.