o4-mini Token Counter
Exact token count and cost for o4-mini, worked out in your browser.
The count here is exact
o4-mini is one of the models on this site that can be counted precisely rather than estimated, because a matching tokeniser runs locally in the page. The number you see for a given input is the number the API will bill.
That exactness applies to the input only, and for a reasoning model the input is not where the interesting cost lives.
Reasoning tokens are the thing this page cannot show you
o4-mini thinks before it answers, and that thinking is billed as output at the full output rate. Those tokens are generated by the model, so nothing you paste can predict them.
The practical effect is that the visible answer length is a poor guide to what a call costs. A two-sentence reply can sit on top of thousands of reasoning tokens on a hard problem, and on an easy one it will not. Two prompts of identical length can differ severalfold in cost depending on how much deliberation the model decides they warrant.
Budget from measured usage, not from this page's number plus a guess at the answer length. Run a representative sample through the API and read the reported output tokens.
More detail
The price
$1.10 per million input and $4.40 per million output, a 4 to 1 ratio. That is cheap for a reasoning model and is the reason to reach for it, but the reasoning-token effect means the effective cost per call is higher than the headline suggests on anything genuinely difficult.
For work that needs no deliberation at all, a non-reasoning model at a similar price will be cheaper in practice as well as faster.
Context
200,000 tokens, which is generous for the tasks this model suits. Reasoning models are typically applied to a hard problem stated compactly rather than to an enormous document, so the window is rarely the binding constraint.
Questions
Why does my bill exceed what this page predicted?
Reasoning tokens. The model generates them before answering and they are billed as output at the full rate. They are invisible in the reply and impossible to predict from the input, so measure real usage rather than estimating.
Can I see how many reasoning tokens a call used?
Not from here. The API response reports its own output token usage, which includes them. That is the number to budget from.
Is the input count exact?
Yes. A matching tokeniser runs locally for this model, so the input figure is precise. It is the output side that cannot be predicted in advance.
When should I not use a reasoning model?
When the task needs no deliberation: formatting, extraction, classification, simple rewriting. You pay for thinking that adds nothing, and you wait longer for it.