The cost of reasoning: what thinking models charge for output

A FinOps guide by OptimNow · prices from the live OptimToken catalogue

A reasoning control used to be something you shopped for. It is now standard equipment. Of the 288 models in the OptimToken catalogue, 213 expose one, which means the provider accepts a thinking budget or an effort setting on the request. The shortlist question has moved with it. What matters is no longer which models can think, but what the thinking costs on the model you already picked. That answer sits on the output side of the rate card.

Reasoning tokens are output tokens

A model that reasons before answering generates the reasoning. Those tokens are billed at the output rate whether or not the API returns them to you. They also never appear in the prompt you wrote, so an estimate built from prompt length misses them completely.

The rate card already shows the effect. Output is priced at 5x input or more on 47% of the 207 reasoning-capable models, against 20% of the 74 that carry no reasoning control. At the steep end the split is sharper: of the 35 models pricing output at 8x input or more, 33 expose a reasoning control.

What a thinking budget costs

Take the Knowledge Q&A preset from the pricing table, at 2,000 input tokens and 800 output tokens per request. Add 1,000 thinking tokens, a round budget chosen for the arithmetic rather than measured from any workload. The median reasoning-capable model then costs 78% more per request, on an identical prompt and an identical answer back to the user.

What sets the size of that increase is the output-to-input ratio. Here is the highest Arena-ranked reasoning model in each ratio band, priced on the same request:

The uplift follows the ratio and not the price level. Claude Opus 4.6 charges $25.00 per 1M output tokens and takes +83%. Gemini 3.5 Flash Lite charges $2.50 and takes +96%. The cheaper model absorbs a thinking budget worse than the dearer one, which is what a blended input-output price hides.

Put DeepSeek V3.2 and Gemini 3.5 Flash Lite side by side if your workload is reasoning-heavy and the quality gap is one you can measure.

The discounts do not follow the tokens

Prompt caching discounts input tokens, and a thinking budget spends output tokens. The two levers sit on different halves of the bill, and the half that grows under reasoning is the one caching cannot reach.

Take the same Knowledge Q&A request, across the 156 reasoning-capable models whose published cache-read rate lowers the figure. Caching removes a median 21.0% of the bill with no thinking budget. At 1,000 thinking tokens it removes 11.5%.

That decides how the number should be reported. A team quoting its caching programme as a percentage will watch the percentage fall while the programme works as designed. The budget at which the share halves is the list cost of the request expressed in output tokens. It is a per-model figure, at a median of 1,300 thinking tokens, and it sits under 2,000 for 155 of the 156 models measured.

One published discount does reach an output token, and it is the batch tier. Of the 207 reasoning-capable models, 56 publish a batch output rate, at a median 50% off list. The constraint is the one batch always carries: nobody can be waiting on the answer. That rules out the interactive shapes, agent loops, coding assistants and tool-calling chains, whatever their thinking budget. On scheduled reasoning work, such as document classification with a verification pass or an evaluation suite, it halves the thinking bill and the mechanics are here.

Where this does not apply

The argument assumes output costs more than input, which is not universal. DeepSeek V3.2 prices output at 1.49x input, so 1,000 thinking tokens add 47% to the request rather than the 78% the median model takes, and it is the highest Arena-ranked reasoning model priced under 2x. At the flat end of the catalogue, 2 models price a thinking token at exactly the rate of a prompt token (Nemotron 3.5 Content Safety and R1 Distill Llama 70B). On those rows a thinking budget costs the same as the equivalent number of prompt tokens, and the output rate decides nothing. Check the ratio first.

The tag itself needs one caveat. It records that the provider accepts a reasoning parameter, read from the upstream catalogue. It does not say thinking is on by default, how many tokens a given effort setting spends, or that the model is any good at reasoning. Until 2026-09-21 this site inferred the tag from the model's name, which tagged an image generator, a vendor called Thinking Machines and a model whose name contains "pro4". The name rule found 46 models on that day. The parameter finds 213 today. That correction is what made this page possible, and it is also why the population is broad enough that the tag is a weak filter now.

What to measure

Three checks, in the order they pay off. Set an explicit thinking budget or effort level on every call. An unbounded reasoning budget is the token equivalent of an unbounded max_tokens, and the default setting is rarely the cheapest one your accuracy target allows. Then log realised output tokens per request against the estimate the feature was costed on. Reasoning spend can grow with no change to your code and none to your prompt, which makes it the line an estimate built from the request misses. Last, when the shortlist is down to two models, compare them on the output rate rather than a blended one. The median reasoning-capable model prices output at 4.2x its input, so a 30/70 blend already understates the output side of a workload that thinks. Ask for the ratio.

Figures on this page come from the 2026-09-27 price snapshot of 281 priced models and rebuild nightly. Thinking budgets are stated in output tokens rather than estimated from a workload. The hub holds no data on how many reasoning tokens a given model spends, and inventing that number would put a guess underneath every figure above.

The FinOps reasoning behind this guide comes from OptimNow's open-source practice library: github.com/OptimNow/cloud-finops-skills. If you want your own reasoning workloads costed against the output rates you are paying, OptimNow runs that review.

Data snapshot: 2026-09-27 · Compare all models · Cloud compute pricing · All guides