[Feature]: Invoke Claude Thinking when using OpenAI API

[mwl38765]

The Feature

When using the OpenAI API to invoke a model with reasoning/thinking, a "reasoning_effort" parameter is supplied. However, with Claude Sonnet 3.7, it requires a "thinking" parameter.

Understand that it's possible to just configure LiteLLM to invoke thinking on every prompt sent through the litellm_params for the model.

But it would be helpful if the reasoning_effort in OpenAI got translated to a thinking value before it is sent to a service hosting Claude Sonnet 3.7.

Motivation, pitch

Having this feature would allow those using LiteLLM for the OpenAI API access to have the flexibility of accessing the thinking feature of Claude Sonnet. The current workaround of having thinking on for every prompt would be an extra cost when the thinking feature is not needed.

Are you a ML Ops Team?

No

Twitter / LinkedIn details

No response

[ishaan-jaff]

For the v0 we should support this on Anthropic / Bedrock Claude models

[ishaan-jaff]

note: needs to be added for Anthropic, Vertex Anthropic, Bedrock Converse Anthropic and Bedrock Invoke Anthropic