o3 Token Counter
Exact token count and cost for o3, worked out in your browser.
An exact input count for a model whose output you cannot predict
o3 is one of the few models here that can be tokenised precisely rather than estimated, so the input figure on this page is what the API will bill. That is genuinely useful for checking you fit inside the 200,000-token window.
It is also only half the story, because o3 is a reasoning model and the half you cannot see is where the money goes.
Why the input number is the smaller half
Before answering, o3 generates reasoning tokens. They are billed as output at $8 per million, they do not appear in the response you read, and nothing about your prompt determines how many there will be.
The model decides how much deliberation a problem warrants, so two prompts of identical length can differ by a factor of several in what they cost. A short question that turns out to be hard is more expensive than a long one that turns out to be easy. This is the opposite of the intuition most people bring from non-reasoning models.
Budget from measured API usage, not from this page plus a guess at the reply length.
More detail
Against o4-mini
o4-mini is $1.10 and $4.40 against o3's $2 and $8, so roughly half the price with the same 200K window and the same 4 to 1 ratio. Both are reasoning models and both are counted exactly here.
The choice is capability, not economics. o3 is the heavier reasoner and earns its price on problems where the extra deliberation changes the answer. On problems where it does not, you are paying twice as much for the same result more slowly. Run a representative sample through both and compare outputs before defaulting to the larger one.
The window and the reasoning that competes for it
200,000 tokens, and the reasoning tokens consume part of it as they are produced. On a hard problem with a long chain of thought, the room actually available for your input is smaller than the raw number suggests, so do not plan to fill it.
Questions
Is the input count for o3 exact?
Yes. A matching tokeniser runs locally in the page, so the input figure is precise rather than estimated. The output side is what cannot be known in advance.
Why did a short prompt cost so much?
Because o3 decided the problem was hard and generated a long chain of reasoning, billed as output at $8 per million. Difficulty drives the cost far more than prompt length does.
Should I use o3 or o4-mini?
o4-mini at half the price is the right default, and o3 is worth it only where the extra deliberation demonstrably changes the answer. Test both on a real sample rather than assuming the bigger model is better for your task.
Can I stop it reasoning so much?
Not from the prompt in any reliable way. If a task genuinely needs no deliberation, the answer is a non-reasoning model, which will be cheaper and faster for that work.