Pydantic AI 2.67 moves native-tool options

Pydantic AI released 2.67.0 on November 10, 2026. The release notes open with 22 compatibility notes, so read this one before you bump the pin, especially if you use native provider tools, RepoContext or durable execution.

Native tools take provider_settings

Provider-specific fields on native tools now live under provider_settings. That is also how the release adds Anthropic's web_search_20260318 and web_fetch_20260318 tools. If you set provider-only options on a native tool today, expect to move them.

FileSearchTool.max_num_results is now actually sent to OpenAI Responses, and openai_service_tier accepts the fast and ultrafast tiers.

RepoContext loads instructions differently

RepoContext now loads one instruction file per directory, includes nested files by default, and gains a walk_up option. Inlined nested instructions are capped at 32,000 characters per conversation by default, adjustable with nested_max_chars. Nested instruction files also surface when the model writes or edits in their directory. Agents that relied on the old loading order may see different prompts.

Durable execution covers more steps

• Input, output and tool guardrail verdicts are recorded as durable operations.
• SubAgents(dynamic_agents=...) and SubAgents(task_control=...) work under durable execution.
• DBOSDurability can run tool calls as DBOS child workflows, in parallel when a workspace is attached.
• A StepPersistence tool call that ends in ModelRetry or ToolFailed is recorded as failed instead of staying started.

Smaller changes worth knowing

CodeMode rejects tools named after Python builtins, which the sandbox used to call instead of your tool, and its remote sandboxes now run over gRPC with pydantic-monty 1.5.0. pydantic-ai-harness swaps httpx for httpx2. A queued message can be changed with dequeue and set_enqueue_priority before delivery. New model names include Claude Fable 5.2, Claude Mythos 5.2 and Gemini 4 Carbon, and KnownModelName has been trimmed to models a user would pick, which can break type checks on removed literals.

This entry reports the documented changes; we did not benchmark the release.

See the Pydantic AI hub, the earlier Pydantic AI 2.64 note, or compare options in Pydantic AI alternatives.