Thinking tokens are billed at the output rate and hidden from the response. A p50 (typical) call and a p99 (hard) call on the same model run identical code but land far apart on the bill. The gap — not the average — is what wrecks a monthly budget.
Naive estimate vs typical vs tail
The "naive" row is what a plain token calculator tells you — it ignores hidden thinking entirely. The gap to the typical row is the thinking premium you'll actually pay; the gap to the tail row is the budget risk.
| Scenario | Thinking tokens | Cost / call | Cost / month | vs naive |
|---|
Budget band by tail severity
How wide your monthly bill can swing depending on how heavy the hard-prompt tail gets. Pick the tail multiple that matches how variable your prompts really are.
| Tail multiple | p99 thinking tokens | Cost / call | Cost / month |
|---|
Why a single cost-per-call number lies about reasoning models
A non-reasoning model is close to deterministic on cost: you send N input tokens, you get back M output tokens, you multiply by the rates and you're done. Reasoning models break that. Between the prompt and the answer the model runs a private chain of thought — the thinking tokens — and you are billed for every one of them at the output rate even though they never appear in the response. Crucially, the count is data-dependent: an easy classification might spend a few hundred, a gnarly multi-step proof or an agentic tool-planning turn might spend tens of thousands. So the true cost per call isn't a number, it's a distribution, and quoting only its median (p50) hides the part of the distribution that actually blows budgets: the tail.
The p50 / p99 spread and the reasoning tax
Cost per call is input×input_price + (visible_output + thinking)×output_price. Hold everything fixed except thinking and the whole thing pivots on that one variable. At the typical thinking count you get your p50 cost; multiply thinking by a tail factor for the hard-prompt case and you get p99. With a heavy reasoner the two can sit 3–5× apart, which means the same monthly volume can bill anywhere across that band depending only on how hard the month's prompts happened to be. The reasoning tax — the fraction of the bill that is thinking rather than the input you sent or the answer you received — routinely clears 70–80% on hard workloads. You are paying mostly for scratch-work. Seeing that share is usually the moment teams add a thinking cap.
How to shrink the band
Three levers move it. Cap the reasoning effort or max-thinking-tokens so the tail physically can't run away — this trades a little answer quality on the hardest prompts for a hard ceiling on cost. Route by difficulty: send the easy majority to a cheap non-reasoning model and reserve the reasoning model for prompts that genuinely need it. And monitor the p99, not the mean, because a batch of hard prompts raises the tail long before it moves the average. Note that prompt caching cuts the input side but not the thinking side, so on reasoning-heavy traffic the thinking cap is the lever that matters — pair this with the flat reasoning token calculator for the point estimate and the agent loop budget calculator for the multi-step case.
How this calculator works
It prices a reasoning call as input×input_price + (visible_output + thinking)×output_price per million tokens, computing three points: a naive estimate that ignores thinking, a typical (p50) call at your typical thinking count, and a tail (p99) call at p50 thinking × your tail multiple. It reports the reasoning tax (thinking's share of the p50 bill), the typical and tail monthly totals at your call volume, and a sensitivity table across tail multiples so you can see the full budget band.
Frequently asked questions
Why are reasoning models so unpredictable to budget?
They emit hidden thinking tokens billed as output, and the count swings from a few hundred on an easy prompt to tens of thousands on a hard one. Size the budget on the typical call and the tail can bill 3–5× more per call.
How do I calculate reasoning token cost?
Cost per call = input×input_price + (visible_output + thinking)×output_price, per million tokens. Thinking is billed at the output rate and is invisible and variable — estimate a typical value and a tail multiple and price both.
What is the reasoning tax?
The share of your bill that is hidden thinking rather than the input or the answer. On heavy-reasoning calls it routinely exceeds 70–80% — you pay more for scratch-work than for the visible conversation.