Pydantic AI AI Updates: November 3, 2026

1. Pydantic AI 2.65.0 Adds a ConversationStore, OpenAI Responses Over WebSocket and Fallback on Refusals

Pydantic AI released v2.65.0 on November 2. New features include a WebSocket transport for the OpenAI Responses API, a ConversationStore that keeps a conversation as a log of messages each stored once, with Conversation.revision to catch conflicting saves, ModelFallback moving to the next model when a model refuses (with ContentFilterError in fallback_on now matching refusals), the '12h' prompt cache retention on the Responses API for GPT-5.6 and later, native mid-conversation system prompts on OpenRouter’s OpenAI routes and on Claude models that support them, and an anthropic_web_tools_allowed_callers setting; importing a name removed in V2, such as OpenAIModel or MCPServerStdio, now names its replacement. Compatibility notes: Z.ai balance and plan-limit 429s raise ModelQuotaExceededError instead of ModelRateLimitError, Google requests get a server deadline of at least 20 minutes rather than the client’s read timeout (with a warning about httpx’s 5 second default), Temporal by default no longer retries model errors that cannot succeed on retry (configurable through non_retryable_model_errors), and BrowserUse model calls now count toward the host run’s usage and UsageLimits. Cost-related fixes move the OpenRouter Gemini cache breakpoint before the final message so it no longer doubles the billed prompt and forward cache='1h' to Claude Haiku 5.5 on Bedrock instead of reducing it to '5m'; <context> elements in user and tool content are now marked as not coming from the application, and Coder returns a tool’s error to the model once its retries run out instead of ending the run. Source

2. Pydantic Publishes a Criteria-Based Comparison of Python Agent Frameworks

Pydantic published “The best Python AI agent frameworks in 2026” by Douwe Maan on November 2, a vendor-authored comparison that ranks Pydantic AI first and covers LangGraph and LangChain, the OpenAI Agents SDK, Google ADK, Microsoft Agent Framework, Strands Agents, CrewAI, the Claude Agent SDK and Agno, plus LiveKit Agents and Pipecat for voice, based on each project’s docs and license files as of October 29. Its criteria are switching models without beta adapters, durable execution that survives a process dying mid-run, human approval that survives the process going away, native frontend and voice protocols, OpenTelemetry tracing to any backend, how far the framework can be taken apart, and commercial support for the open-source framework itself. For Pydantic AI it counts 30+ native model providers, eight durable execution engines (Temporal, DBOS, Prefect, Restate, AWS Lambda, Absurd, pgtask and Apache Airflow), ten sandboxes and remote workspaces, and five speech-to-speech providers, and lists its own catches as Python only, no hosted platform or management UI, and fewer retrieval integrations than LangChain (about 80 document loaders, 80 vector stores and 90 retrievers). For the others it records, among other points, that LangGraph relies on its own checkpointers and marks its MCP client beta, that the OpenAI Agents SDK reaches non-OpenAI models through adapters its docs call best-effort and beta and uploads traces to OpenAI by default, that Google ADK’s tool confirmation is experimental, that Strands’ durable execution comes from a Temporal plugin in public preview, and that the Claude Agent SDK runs Claude models only with no documented durable execution engine integration; Pydantic says it re-checks the comparison against each project’s docs every month. Source