How reasoning tokens work

A normal request has two obvious sides: the tokens you send and the output tokens the model generates. A reasoning-capable model can divide its generated output into another functional stage:

  1. It reads the input.
  2. It generates reasoning tokens to work through the task.
  3. It generates the final answer.

That three-step picture is useful, but it is not universal. Some models can reason between tool calls or between pieces of visible output rather than completing one uninterrupted private scratchpad first.

The word reasoning describes what the generated tokens are being used for. It does not mean that other model computation stops, or that ordinary models perform no reasoning at all. Every model performs numerical computation to select its next token. The special feature here is that a model spends generated tokens on intermediate work before or between the tokens intended for the user.

Training can encourage this behavior. For example, reinforcement learning can reward correct final answers and lead a model to develop generated patterns such as checking a result, changing strategy, or revisiting an earlier step. More reasoning tokens therefore mean more inference-time work, not more training or a change to the model's stored knowledge.

Hidden reasoning and visible summaries are different

“Reasoning token” is an umbrella term, not a standard shared by every provider.

OpenAI reports a reasoning-token count but does not expose the raw reasoning through its API. Google can return a summary of its thoughts and an opaque signature representing reasoning state. Anthropic's interfaces and controls vary by model, with some exposing summarized thinking and some using adaptive effort instead of a fixed token budget.

This creates three distinct things:

  • Raw reasoning tokens: the generated intermediate work.
  • A reasoning summary: a shorter account produced for the user or developer.
  • The final answer: the response intended to complete the request.

A summary is not a transcript of every raw reasoning token. Hiding the raw trace also does not mean the tokens were free or absent. To know what a request used, check the API's usage metadata rather than counting the words displayed on screen.

A token-usage example

Suppose a usage report says:

The non-reasoning portion of the generated output is:

1,186 − 1,024 = 162 tokens

The request did not use 1,186 answer tokens plus another 1,024 reasoning tokens. The 1,024 reasoning tokens are already included in the 1,186-token output total.

This distinction explains how a concise answer can have a much larger output count than its visible text suggests. Field names differ across APIs, but look for a reasoning or thinking count nested in the response's usage details.

Why reasoning tokens matter

Cost

Major providers generally charge reasoning or thinking tokens as generated output. A short final answer can therefore cost more than you would estimate from its visible length. Current rates and accounting rules are provider-specific, so use the provider's pricing and usage documentation for an actual cost calculation.

Output and context limits

Reasoning needs room. Providers commonly count reasoning tokens against an output limit, the context budget, or both. If the generation limit is exhausted during reasoning, a request can end before producing a useful visible answer.

This is why a maximum-output setting cannot always be treated as “maximum answer length.” It may need to cover intermediate reasoning as well as the final response. The exact rule varies by API.

Response time

Reasoning tokens are generated work. More of them generally require more sequential generation before the request finishes, which can increase latency. Effort settings and token budgets are therefore controls over a quality, time, and cost tradeoff—not simply style settings.

Task quality

Extra reasoning is most useful when the task rewards intermediate work: multi-step mathematics, code debugging, planning, constraint checking, or tool-based workflows. Simple extraction or formatting may gain little from it.

More is not automatically better. A model can spend many tokens on an unproductive path and still be wrong. Evaluate the final result on the task you care about rather than treating the reasoning-token count as a quality score.

Common misconceptions

“Reasoning tokens are the model's private thoughts”

That phrase is convenient but too literal. Reasoning tokens are model-generated intermediate sequences. They can resemble written problem-solving, but their existence does not establish human-like thought, understanding, or self-awareness.

“If I can see reasoning, I am seeing the raw tokens”

Not necessarily. A product may show a summary or a filtered explanation. Provider documentation distinguishes those displays from the full internal trace used for accounting.

“A 500-token answer uses 500 output tokens”

Only if there are no other generated tokens counted in the output. With reasoning enabled, the usage total may include hidden reasoning as well as the answer.

“The budget tells me exactly how many tokens the model will use”

It depends on the provider and model. A setting may be a hard maximum, a target, or an effort level. Some models adapt their reasoning amount to the task. Treat the returned usage count as the measurement and the request setting as a control.

Where to go next

Read What Are Output Tokens? to understand the broader accounting category that often contains reasoning tokens. Then use Input vs. Output vs. Reasoning Tokens when you need to compare all three categories in one request.