Pydantic AI AI Updates: October 31, 2026

1. Pydantic AI 2.64.0 Adds ModelFallback, a First-Event Timeout, Artifact Scanning and a Boxd Sandbox

Pydantic AI released v2.64.0 on October 30. A new ModelFallback capability, backed by a per-attempt prepare_model_request hook and a RetryModelRequest exception, retries a request on another model, can be built from agent specs, and takes a sticky option that keeps the rest of a conversation on the fallback model that answered; a companion first_event_timeout model setting lets a stream that never starts fall back, and errors a retrying transport gives up on are now raised as ModelHTTPError so fallback proceeds. ArtifactScanner adds security scanning for Memory and Skills artifacts, BoxdSandbox runs Coder, Shell, FileSystem and other harness capabilities in a boxd machine, and RunContext.load_capability loads on-demand capabilities from code. Other additions include output_speech_timestamps for word timing on Azure AI Voice Live and ElevenLabs Agents, @agent.system_prompt(dynamic=True) with on_change='append' so changes do not invalidate the cached prompt, ModelResponse.model_id recording which model served each request, and Pydantic Evals spans that accept plain functions as evaluators and record CI provenance through OpenTelemetry vcs.* and cicd.* attributes. Compatibility notes: json-repair is now optional and needed only by Coder and RepairToolArguments, Coder tools use non-strict schemas unless Coder(strict=True) is set, stored response, conversation and container IDs are sent only to the endpoint that created them, and RepoContext surfaces every nested instruction file between a traversed path and the working directory. Fixes cover Bedrock native output on older botocore, Gemini Live streams losing a tool call on resume, AnthropicModel.count_tokens undercounting server tools, SubAgents and DynamicWorkflow children now running under the parent’s usage_limits, and web_fetch no longer leaking script and style source into markdown. Source

2. Pydantic Previews Validation V3 as a Low-Disruption Upgrade and Ends V1 Support

Pydantic published a post by Victorien Plot announcing the start of work on Pydantic Validation V3. Breaking changes will be limited to correctness fixes that affect a minority of users or that can be easily reverted to the old behavior, V1 deprecations and the pydantic.v1 namespace will be removed, and the team says there will be no API overhaul on the scale of the V1 to V2 migration; candidate issues are tracked under the v3 label on GitHub, and the goal is to have V3 close to its final state by the end of 2026, with no release date set. Pydantic V1.10.27 is the final V1 release, carrying security fixes that are also included in the pydantic.v1 namespace in V2.14.2, and V1 support has ended, so future security issues in V1 will not be fixed. V2 will keep receiving fixes for major bugs and security issues after V3 ships, with no end date, until most users have moved. Since Pydantic AI, FastAPI and much of the Python agent stack sit on Pydantic Validation, any dependency still importing pydantic.v1 will need to migrate before V3. Source