Reasoning response support

[samuelcolvin]

See:

• OpenAI reasoning docs
• Deepseek reasoning docs
• Gemini "thinking" docs

No idea yet how this should look, but we should try to support it.

[Wh1isper]

I've used some pretty tricky ways to implement reasoning in the pydantic-ai-bedrock package: ai-zerolab/pydantic-ai-bedrock#15

Based on this practice, I think we need to add a ReasoningPart, and then the Model will implement the object transformation based on that. On top of that, we need to think about how the Agent is returned, perhaps using the <Thinking> tag would be better?

[soichisumi]

I would like to request support for Claude's Extended thinking.
While it may be difficult to be compatible with other models, this feature provides robust and flexible thinking and integration with tool use.
https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking?q=thinking#how-extended-thinking-works

[arty-hlr]

Maybe openrouter's reasoning tokens should also be added: https://openrouter.ai/docs/use-cases/reasoning-tokens for people who use Claude 3.7 through it for example. At the moment no reasoning is shown.

[arty-hlr]

Link to relevant slack conversation:
https://pydanticlogfire.slack.com/archives/C083V7PMHHA/p1743405703872439

[pedroallenrevez]

Any news on this?

[brycedrennan]

I want to use pydantic-ai at work but not being able to use reasoning models is one of the things holding me back.

[aiizloli-ecs]

Any update on this ?

[seangal2]

Hi,

I’d like to express interest in supporting extended reasoning models, and contribute a few observations.

Looking at PR #1142, I noticed that the current implementation returns the summary field from OpenAI as the reasoning part. However, OpenAI’s responses API also supports an optional encrypted_content field (reference, docs), which is useful for including full reasoning traces in follow-up interactions. This is especially recommended when using tool calling (docs).

To support this, I suggest either:

• Extending ThinkingPart to include the encrypted_content field, or
• Introducing a new kind of part specifically for encrypted or provider-specific reasoning items.

It’s worth noting that encrypted_content is provider-dependent and may not be interoperable - for example, it can’t be mapped to Anthropic’s redacted_thinking. Still, a unified abstraction might help represent these different formats consistently.

Happy to help further if there’s interest!