AI token calculator
Count the tokens in a prompt, set how long the answer will be, and see what the same call costs on 25 models from Anthropic, OpenAI and Google. No sign up.
Expected output size
The same call, priced on every model.
Cheapest first, inside each provider. Move the output slider above and the order changes.
Cheapest at these settings: calculating
Most expensive: calculating
Anthropic
Claude models, 7 priced
- Claude Haiku 4.5Small and fast $1.00 / $5.00 $18.00 $0.0180 each
- Claude Sonnet 5Balanced * $3.00 / $15.00 $54.00 $0.0540 each
- Claude Sonnet 4.6Balanced $3.00 / $15.00 $54.00 $0.0540 each
- Claude Opus 5Frontier $5.00 / $25.00 $90.00 $0.0900 each
- Claude Opus 4.8Frontier $5.00 / $25.00 $90.00 $0.0900 each
- Claude Opus 4.7Frontier $5.00 / $25.00 $90.00 $0.0900 each
- Claude Fable 5Frontier $10.00 / $50.00 $180.00 $0.1800 each
Prompt caching reads bill at roughly a tenth of the input rate. Batch requests bill at half.
OpenAI
GPT models, 10 priced
- GPT-5.6 LunaSmall and fast $0.20 / $1.20 $4.08 $0.00408 each
- GPT-5.4 nanoSmall and fast $0.20 / $1.25 $4.20 $0.0042 each
- GPT-5.4 miniSmall and fast $0.75 / $4.50 $15.30 $0.0153 each
- GPT-5.6 TerraBalanced $2.00 / $12.00 $40.80 $0.0408 each
- GPT-5.2Balanced $1.75 / $14.00 $44.10 $0.0441 each
- GPT-5.4Balanced $2.50 / $15.00 $51.00 $0.0510 each
- GPT-5.6 SolFrontier $5.00 / $30.00 $102.00 $0.1020 each
- GPT-5.5Frontier $5.00 / $30.00 $102.00 $0.1020 each
- GPT-5.5 ProFrontier $30.00 / $180.00 $612.00 $0.6120 each
- GPT-5.4 ProFrontier $30.00 / $180.00 $612.00 $0.6120 each
Cached input bills at a tenth of the input rate on most models. Batch requests bill at half.
Google Gemini
Gemini models, 8 priced
- Gemini 3.1 Flash-LiteSmall and fast $0.25 / $1.50 $5.10 $0.0051 each
- Gemini 3.5 Flash-LiteSmall and fast $0.30 / $2.50 $7.80 $0.0078 each
- Gemini 2.5 FlashSmall and fast $0.30 / $2.50 $7.80 $0.0078 each
- Gemini 3.7 FlashBalanced * $0.75 / $3.75 $13.50 $0.0135 each
- Gemini 3.6 FlashBalanced * $0.75 / $3.75 $13.50 $0.0135 each
- Gemini 3.5 FlashBalanced $1.50 / $9.00 $30.60 $0.0306 each
- Gemini 2.5 ProFrontier * $1.25 / $10.00 $31.50 $0.0315 each
- Gemini 3.1 ProFrontier * $2.00 / $12.00 $40.80 $0.0408 each
Pro models carry a second, higher price band above 200,000 input tokens. Batch requests bill at half.
Published on-demand rates per million tokens, last checked 2026-08-20. Batch, cached input and committed-use rates are lower.
* Claude Sonnet 5: Introductory pricing of $2.00 and $10.00 runs through August 31, 2026. Gemini 3.1 Pro: Preview pricing. Above 200,000 input tokens the rate rises to $4.00 and $18.00. Gemini 2.5 Pro: Above 200,000 input tokens the rate rises to $2.50 and $15.00. Gemini 3.7 Flash: Promotional rate through December 31, 2026. It doubles on January 1, 2027. Gemini 3.6 Flash: Promotional rate through December 31, 2026. It doubles on January 1, 2027.
How do you work out what an AI call costs?
Multiply your input tokens by the model's input rate, multiply the tokens it writes back by the output rate, and add the two together. Both rates are quoted per million tokens, so a 6,000 token prompt returning 300 tokens on a model priced at $5.00 and $25.00 costs $0.0375 per call, or $37.50 across a thousand calls. Output is the half that moves: it is three to six times the input rate on almost every model, so the length of the answer changes the ranking more than the length of the prompt does.
Four things the number above is telling you.
- Output length decides the winner
- At a 5% output ratio the cheapest model on this page is a different one than at 200%. Set the slider to your real traffic shape before comparing anything.
- The spread between tiers is very wide
- Across these 25 models the top of the range costs more than a hundred times the bottom for the identical call. Most production traffic is classification, extraction and routing, where the bottom of the range scores the same on your own evals.
- List price is rarely the paid price
- Prompt caching, batch pricing and committed-use tiers each move the real figure. A model that looks second cheapest on list price is often first once caching is applied to a stable system prompt.
- One call is not the question
- The figure that matters is calls per month times cost per call, split by the workflow that made them. That is the number this page cannot give you, because it lives in your billing data.
This page prices a call. Your invoice prices a month.
A calculator answers what one call should cost. It cannot tell you how many calls you made, which team made them, which feature they served, or which of them are running on a model that costs more than the job needs.
Tokens, pricing and this calculator
What is a token?
A token is the unit every model bills in. It is a fragment of text, usually three to four characters of English, so a short word is often one token and a longer word is two or three. Punctuation and spaces count too.
You pay for tokens twice in every call: once for everything you send, which is the input, and once for everything the model writes back, which is the output. Output is the expensive half on nearly every model on this page.
Do Claude, GPT and Gemini count the same text as the same number of tokens?
No, and it is worth knowing by how much. Every provider ships its own tokenizer, and some change it between model generations. For ordinary English prose the counts land close together, but they are not identical, and on code, JSON, non-Latin scripts and heavy punctuation they separate further. Anthropic's newer Claude models tokenize more finely than earlier ones, so the same text can register a higher token count there than the estimate on this page.
This page deliberately applies one estimate to all three providers. That keeps the comparison honest about the thing it is actually comparing, which is the price per token. Treat the ranking as sound and the absolute figure as an approximation.
For an exact count, each provider offers a token counting endpoint that returns the real number for a specific model before you send the request. Use it on a real sample of your traffic if you are sizing a budget rather than sizing a decision.
Why does output size change the ranking?
Because output costs three to six times what input costs, and the multiple is different on every model. A model that looks cheap on a classification job, where the answer is a few tokens, can be the expensive option on a long generation job.
That is what the slider is for. Set it to the shape of your real traffic and the order of the list changes.
Are these the prices we would actually pay?
They are the published on-demand rates, which is what most companies pay before anything is negotiated. Three things routinely bring the real figure down: prompt caching, which bills repeated context at roughly a tenth of the input rate, batch processing at half price for work that can wait, and committed-use tiers once volume is predictable.
Those three are often worth more than switching provider. Cost optimization covers how each one is found.
How current is this list?
Prices were last checked on 2026-08-20. The model market reprices often, so treat any figure older than a month as approximate and confirm it against the provider's own price page before you build a plan on it.
Does price alone tell us which model to use?
No. Price per token only means something next to quality on your own task and the number of tokens the model needs to finish it. A model at half the rate that takes three attempts is not the cheaper option.
Run your evals first, then use this page to price the models that passed.