Skip to content

Reasoning Token Cost Estimator

What hidden thinking tokens add to a bill, what turning effort down would save, and what share of the invoice is text you never see.

Monthly spend with reasoning on
$2,160

$1,350 of that is thinking you never see — 2.67× the bill with reasoning off.

Billed output tokens per reasoning request
1,200
…of which is hidden reasoning
75.0%
Cost per request, reasoning on
$0.02
Cost per request, reasoning off
$0.0081
Cost per request, lower effort
$0.01
Monthly, reasoning on
$2,160
Monthly, reasoning off
$810.00
Monthly at the lower effort setting
$1,260
Saved by turning effort down
$900.00
Reasoning tokens as a share of the whole bill
62.5%
What this assumes: reasoning tokens are billed at the output rate — that is how every provider that charges for them does it today, but check your own price list before trusting the figure; they are counted in the output cap, so a long think can truncate the answer you were paying for; the token counts here are yours to supply from a real response's usage object, because hidden reasoning cannot be counted from the text you receive; effort settings change the average, not the ceiling, and a hard question will still think longer than your typical figure; no caching of any kind is modelled.

A reasoning model's bill does not look like its output. You see three hundred tokens of answer; you are invoiced for that plus however much the model thought first, at the output rate, which is typically several times the input rate. That is the whole arithmetic on this page, and the row worth reading twice is "hidden reasoning as a share of billed output" — when it is above two-thirds, most of what you are buying is text nobody will ever read.

Which is not an argument against it. The comparison that matters is not reasoning versus no reasoning at the same model, it is reasoning versus whatever you would otherwise do to get the same accuracy: a larger model, a chain-of-thought prompt whose tokens you would pay for visibly, three sampled attempts with a judge, or a human reviewing the output. Reasoning is often the cheapest of those. It only looks expensive because the cost arrives on a line you did not previously have.

The practical lever is effort, and the fields above are arranged so you can price it: run one representative batch at each setting, read the reasoning token counts out of the usage objects, and put both numbers in. Then decide with the difference in front of you rather than by feel. What this does not capture: the variance, which on reasoning models is severe — the same prompt can think for 200 tokens or 4,000 depending on how hard it finds the question — and the latency cost of waiting for it, which is a product decision rather than a billing one.

Reasoning Token Cost Estimator · Multigrid