What reasoning tokens are
A reasoning model does not answer in one pass. It first generates a stream of working tokens — Anthropic describes the model restating what is asked, trying approaches, checking intermediate results and abandoning paths that do not hold up — and only then writes the response. Those working tokens are the reasoning tokens. They are what makes these models better at maths, code and long multi-step tasks, and they are also why the same question costs more on a reasoning model.
You rarely see them. OpenAI does not expose raw reasoning tokens at all and offers an optional summary; Anthropic returns either a summary of Claude's thinking or nothing, never the raw chain; Google returns optional thought summaries. In every case the summary is not what you are billed for.
How reasoning tokens are billed
All three major vendors bill reasoning as output. OpenAI states that reasoning tokens occupy space in the context window and are billed as output tokens. Anthropic states that the tokens Claude spends thinking are billed as output tokens even when the thinking text is not returned, and that with summarized thinking you are charged for the full thinking tokens, not the summary — the billed output count will not match the tokens you see. Google says response pricing is the sum of output and thinking tokens, and that the price reflects the full thought tokens the model generates even though only summaries come back.
Usage reports show them, if you look. Google reports total_thought_tokens; OpenAI breaks reasoning tokens out within output usage. On Anthropic models that keep previous thinking blocks, those blocks are later resent as part of the conversation and bill again as input on following turns.
Worked exampleThe same answer, three amounts of thinking
10,000 requests a month on GPT-6 Sol, priced at its current list rates, checked on 2026-09-30.
The reasoning volumes are illustrations inside the range OpenAI describes — a few hundred to tens of thousands per response. The visible answer is 500 tokens in every column; everything above that line is thinking you pay output rates for.
Change the workload or the model in the LLM API cost calculator.
Controlling how many you pay for
Two dials, one soft and one hard. The soft dial is reasoning effort — OpenAI's reasoning.effort takes values up to xhigh and max depending on the model, Anthropic's effort runs from low to max, and Google's thinking_level from minimal to high. Lower effort means fewer reasoning tokens on average. Anthropic recommends lowering effort first when a thinking workload costs too much.
The hard dial is the output cap. Reasoning counts toward it everywhere: Anthropic's thinking tokens are a subset of max_tokens, and OpenAI's max_output_tokens limits reasoning, visible output and formatting tokens together. Hit the cap and the response is cut off — at OpenAI, possibly before any visible text — but the reasoning already generated is still billed.