Reasoning Tokens For Thinking Models

OpenRouter introduced a way for applications to receive a supported model's reasoning text separately from its final answer. Its January 2025 announcement showed an API option that placed this text in a separate response field, initially for DeepSeek R1 and models derived from it, with Gemini thinking models listed as upcoming.

That separation lets an application display the intermediate text or pass it into another model's prompt. The announcement illustrated the latter with a toy example that feeds R1's reasoning to a much simpler model, and OpenRouter itself calls the example's utility questionable. The post establishes access to an additional model output, not measured evidence that passing reasoning between models improves a real workflow.