TL;DR
Special non-word meta-tokens generated internally by reasoning models to perform chain-of-thought analysis before producing a final answer.
Unlike traditional output tokens that represent human language, reasoning tokens function as an internal workspace for a model during inference. They are optimized through reinforcement learning to explore hypotheses, verify intermediate steps, and backtrack when a logic path fails. This dynamic inference-time compute scaling decouples the reasoning capability of the system from its static pre-trained parameter count.
Why this matters for your business
They provide developers with a controllable thinking budget, allowing applications to trade computation cost and latency for higher accuracy on complex logic tasks.
