fix(openai): add support for reasoning content in OpenAI chat responses from OpenRouter

[sinanuozdemir]

Description: I added reasoning_content to additional_kwargs when the reasoning field is present in the response message. This enables support for reasoning content from providers like OpenRouter that include this field in their responses.
Issue: #32981
Dependencies: N/A

P.S. this is my first time ever attempting to contribute to an open source package so feel free to rip me to shreds.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Preview	Comments	Updated (UTC)
langchain	Ignored	Preview		Sep 18, 2025 6:32pm

[codspeed-hq]

CodSpeed Instrumentation Performance Report

Merging #32982 will not alter performance

_{Comparing sinanuozdemir:feature/add-reasoning-content-support (741c518) with master (f158cea)}

Summary

✅ 6 untouched
⏩ 8 skipped¹

Footnotes

1. 8 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[ccurme]

Hi @sinanuozdemir, thanks for this.

You should be able to resolve your issue immediately by using ChatDeepSeek (I see from the originating issue you have langchain-deepseek already installed). That model basically subclasses BaseChatOpenAI and extends it with this feature. Let me know if that doesn't work for you.

There is a lot of confusion around this, and it's arguably weird to force use of ChatDeepSeek for some chat completions use cases, so I'm not opposed to going with this PR. If you want to add unit tests and extend to the streaming case, I think we could push it through.

[MaciejoG]

Came across this PR while researching why using the ChatOpenAI with a model from Anthropic does not output any reasoning tokens (for my usecase, I cannot use the ChatAnthropic class).
Slightly modifying the changes from this PR (instead of reasoning look for reasoning_content) yields a positive result!

if hasattr(message, "reasoning_content"):
                generations[0].message.additional_kwargs[
                    "reasoning_content"
                ] = message.reasoning_content

Output contains reasoning_content

content='The capital of France is Paris.' additional_kwargs={'refusal': None, 'reasoning_content': "This is a straightforward factual question. The capital of France is Paris. This is basic geography knowledge that I'm confident about."}

@sinanuozdemir are you going to continue working on this PR? Would be great to make it functional for reasoning_content too!

[sinanuozdemir]

I'm going to try to take this over the finish line in the next few days @ccurme and @MaciejoG! Thank you both for the tips.

[sinanuozdemir]

@ccurme I think I'm ready for another review!

[Saryazdi-Saman]

I came across this PR while running into the same issue. I had been using the OpenAI SDK directly for Claude via OpenRouter, since I couldn’t find documentation that ChatDeepSeek could be used in that context.

"There is a lot of confusion around this, and it's arguably weird to force use of ChatDeepSeek for some chat completions use cases … If you want to add unit tests and extend to the streaming case, I think we could push it through."

@ccurme I agree with your point about developer confusion — requiring ChatDeepSeek for some chat completion use cases isn’t intuitive, especially since OpenRouter's docs direct to ChatOpenAI for Langchain implementation.

What do you think about the idea of inheriting from BaseChatOpenAI and creating a dedicated ChatOpenRouter wrapper? That might make the integration cleaner and reduce confusion. I’d be happy to contribute code for this (in TS and possibly Python) if that direction aligns with the project.