Pydantic AI v2.65.0 (2026-11-02)
Pydantic AI v2.65.0 introduces WebSocket transport for OpenAI Responses, a new log-based ConversationStore with revision tracking, 12-hour prompt cache retention support, and improved ModelFallback handling on refusals.
原始内容为英文；当前页面提供中文导航与来源说明，具体事实请以原文为准。
为什么值得关注
Pydantic AI release v2.65.0 delivers several feature expansions, compatibility updates, and error-handling refinements across multiple model providers and integrations. Key features include: - Added WebSocket transport to OpenAI Responses. - Introduced ConversationStore , maintaining conversations as single-stored message logs, accompanied by Conversation.revision for conflict detection. - Added multimodal notes to AskUserAnswer . - Added support for allowed callers on Anthropic's web tools ( anthropic web tools allowed callers ). - Enabled '12h' prompt cache retention on the OpenAI Responses API for GPT-5.6 and newer. - Native support for mid-conversation system prompts on OpenRouter OpenAI routes and compatible Claude models. - Enhanced migration feedback by explicitly naming V2 replacements when importing deprecated names like OpenAIModel or MCPServerStdio . Compatibility and Bug Fixes: - ModelFallback now advances to subsequent models when a refusal or ContentFilterError occurs. - BrowserUse model invocations now increment the host run's usage and count against UsageLimits . - Temporal activities are prevented from retrying non-retryable model error subclasses by default. - Bedrock errors on AsyncAnthropicBedrock now raise ModelContextWindowExceededError and ModelQuotaExceededError instead of generic errors. - Z.ai 429 balance and plan limits now surface as ModelQuotaExceededError rather than ModelRateLimitError . - Google model calls now enforce a server deadline of at least 20 minutes, with warnings regarding httpx's 5s read timeout defaults. - Daytona dependency…
可执行摘要
Pydantic AI v2.65.0 adds WebSocket transport for OpenAI Responses, ConversationStore with optimistic locking (Conversation.revision), multimodal notes to AskUserAnswer, and integrates BrowserUse calls into host UsageLimits. Error handling is updated: ModelFallback now steps to subsequent models on refusal/ContentFilterError, Temporal stops retrying unrecoverable model errors, and specific error types…
- Agent 实用度
- 77/100
- 可信度
- 96%
- 机器格式
- JSON + Markdown
开发者应核对什么
- Verify error-handling logic around ModelFallback, as model refusals and ContentFilterError will now trigger fallback models.
- Update code using deprecated V2 imports (such as OpenAIModel or MCPServerStdio) to use their named replacements.
- Ensure Daytona is updated to =0.226.1 if utilizing Daytona integrations.
- Audit Temporal workflows to confirm default non-retryable model error behavior aligns with operational expectations.