Reasoning Token Cost Calculator

Reasoning token cost calculator

✓ Last verified: 2026-07-15· Source: official provider pricing page· Auto-monitored — report change →

Reasoning models think before they answer — and those hidden thinking tokens are billed at the output rate, often 5–15× your visible reply. Enter your input, visible output and reasoning effort, and this shows the true cost per call and per month, how much of your bill is invisible thinking, and what a lighter effort setting would save.

Why your reasoning-model bill is bigger than it looks

A normal chat model charges you for the tokens you send (input) and the tokens it writes back (output). A reasoning model adds a third, invisible category: the thinking tokens it generates internally to work through a problem before composing the answer you see. You never receive that scratch-work, but you pay for it — providers meter it at the output rate, the most expensive tier. The result is a bill that can be several times what a naïve "input + visible output" estimate predicts, because on a hard prompt the model might burn 5,000 tokens of reasoning to produce a 400-token answer.

This calculator makes the hidden half visible. Enter the input and the visible output you expect, choose a reasoning effort (or type the exact reasoning-token count from the API's usage field), and it prices all three streams. The "true monthly cost" includes thinking; the "if you ignored reasoning" figure is the trap — the number you'd get from a token counter that only sees the prompt and the reply. The gap between them is what surprises teams at the end of the month.

The three levers on reasoning cost

1. Effort. Most reasoning APIs expose a low / medium / high (or a token-budget) control. Dropping from high to medium on prompts that don't need deep thought can halve the bill with no visible change in answer quality — the table above shows the exact difference for your workload. 2. Routing. Reasoning is worth paying for on genuinely hard tasks and pure waste on easy ones; send simple requests to a cheaper non-reasoning model and reserve thinking for the hard tail. Size the split with the model routing savings calculator. 3. Prompt design. Clearer, more constrained prompts need less exploration, so they generate fewer reasoning tokens — a rare case where better prompts are also cheaper prompts.

How to use it

1. Pick a reasoning model and enter your monthly request volume.
2. Enter typical input and visible output tokens per request.
3. Choose a reasoning effort, or select "Manual" and type the reasoning tokens you've measured.
4. Read the true cost, see how much is hidden thinking, and use the table to find the effort level that fits your budget.

Common mistakes

Budgeting on visible output only. The thinking tokens are usually the larger half — ignore them and your forecast is wrong by multiples. Leaving effort on high everywhere. High effort is a default, not a requirement; most prompts don't need it. Comparing a reasoning model to a chat model on headline price. Same per-token rate, very different token counts — compare on true cost per request, not per million. Assuming caching helps the reasoning. Prompt caching discounts repeated input; it does nothing for freshly generated reasoning tokens.

FAQ

Do I get charged for tokens I can't even read?

Yes. Reasoning models hide the thinking tokens from the response but still bill them, at the output rate. The API returns the count in a usage field so you can audit it.

How many reasoning tokens is "normal"?

It scales with effort and difficulty — roughly 2× visible output on low effort, ~5× on medium, and 10–15×+ on high effort for hard problems. Measure your own traffic to get an exact multiplier.

Is a reasoning model ever cheaper than a chat model?

Rarely on cost alone — it's cheaper on outcome when a hard task would otherwise take a bigger model or several retries. For easy tasks, a non-reasoning model wins on price every time.

Can I cap reasoning spend?

On many APIs, yes — via the reasoning-effort setting or an explicit thinking-token budget. Lowering it is the single biggest lever on this bill; the table shows the saving.

Estimate only. Reasoning-token usage depends on your prompts and the provider's effort control — verify against real API usage data before budgeting.

Frequently asked questions

Are reasoning tokens billed, and at what rate?

Yes. Reasoning models generate hidden 'thinking' tokens before they write the visible answer, and providers bill those thinking tokens at the standard output rate — the same price as the words you actually see. You don't get to read them, but they count on your invoice. On a hard prompt the reasoning can run several times longer than the final answer, so the thinking, not the reply, becomes the bulk of the cost.

How many reasoning tokens does a request use?

It depends on the reasoning effort setting and the difficulty of the prompt. As a rough guide, a low or 'minimal' effort adds a couple of times the visible output in thinking tokens, medium effort five-ish times, and high effort ten to fifteen times or more on genuinely hard problems. This calculator lets you pick an effort preset or type an exact reasoning-token count if you've measured it from the API's usage field.

How do I cut reasoning-token cost?

Lower the reasoning effort for prompts that don't need deep thought, cap the reasoning budget where the API allows it, and route only genuinely hard requests to a reasoning model while everything else goes to a cheaper non-reasoning model. Because reasoning is billed at the output rate, trimming it is often a bigger lever than shortening your prompt — this calculator shows the saving from each effort level side by side.