🤖 AI Token Calculator
Paste a prompt, pick a model, and see the estimated tokens and what the call costs. Prices are taken from each provider's own pricing page and the date they were checked is shown with every result — which matters, because several of them have an increase already scheduled.
Prompt inspector
Padding costs the same as content. This finds the parts you can drop without changing the meaning.
Removes padding and filler, never instructions. Re-read the result before using it.
What a token is, and why nobody can tell you exactly how many you used
A token is a chunk of text, usually a few characters long. Models do not read letters or words; they read these chunks, and every provider charges by how many go in and how many come out. For ordinary English prose a token averages about four characters, so roughly three tokens for every four words — but that average hides a lot.
| Content | Chars per token | 1,000 words is about |
|---|---|---|
| English prose | 4.0 | 750 tokens |
| Code or markup | 3.1 | 970 tokens |
| Spanish, French, German | 3.0 | 1,000 tokens |
| Chinese, Japanese, Korean | 1.5 | 2,000 tokens |
That last row is the one people are not warned about. The same meaning expressed in Chinese or Japanese can cost two to three times as much as in English, because those scripts tokenize far less efficiently. If you run a multilingual product, your per-user cost varies by language in a way your pricing page probably does not.
Why this is an estimate and not a count
Every provider uses its own tokenizer, and they disagree. The same sentence can be 28 tokens on one model and 31 on another. On top of that, the number you are billed for includes things you never typed: the message formatting the API wraps around your text, the system prompt, any tool or function definitions you have attached, and on reasoning models the internal thinking tokens.
So the figure here is a calibrated estimate, accurate to roughly ten percent for ordinary text, and it is deliberately labelled as one. Any site claiming an exact count across all providers from a single tokenizer is overstating what it knows. For a real number, send one request and read the usage field in the response — that is the only count that gets billed.
Output is the expensive half. Across the models on this page, output costs between three and eight times input. A prompt you agonise over for an hour and a reply you never read cost about the same to produce, and the reply usually costs more.
The cost structure nobody budgets for
Three things reliably blow up an AI bill, and none of them is the prompt. The first is conversation history, because a chat resends everything each turn — the conversation cost calculator shows how fast that compounds. The second is retrieval: attaching documents puts tens of thousands of tokens into every call. The third is retries, where a failed parse quietly doubles the cost of that request.
The defence against all three is the same: cache the parts that do not change, measure the real usage rather than estimating it, and pick the smallest model that passes your quality bar. The spread between the cheapest and dearest model here is a hundredfold on input, which is a bigger lever than any prompt rewrite.
Prices move, and some have a date on them
AI pricing is not stable the way electricity pricing is stable. Models are launched, repriced and retired within months, and some current prices are explicitly promotional with a published end date. Every page in this section shows the date the prices were checked, which was 2026-10-26, and flags any increase that has already been announced. If you are building a budget, build it against a dated figure and recheck it quarterly.
A worked example
A support assistant handles 1,000 conversations a month. Each has a 500-token system prompt, about 150 tokens of user message and a 400-token reply. Treated as single calls that is 650 input and 400 output per conversation: at Sonnet 5.5 rates, $1.30 of input and $4.00 of output, so $5.30 a month. Run as a ten-turn conversation instead, with the history resent each turn, the input alone becomes roughly 33,000 tokens per conversation and the bill is closer to $106. Same product, twenty times the cost, and the only thing that changed was the shape of the interaction.
The natural follow-ups are the llm pricing comparison, for what this figure feeds into, and the ai conversation cost, for the closely related case.
Frequently Asked Questions (FAQs)
How many tokens is 1,000 words?
About 750 for English prose. Code runs closer to 970, and Chinese, Japanese or Korean can reach 2,000 for the same meaning, because those scripts tokenize much less efficiently.
How accurate is this token count?
Within roughly ten percent for ordinary text. It is an estimate from the characters, not the provider's own tokenizer, and it cannot include the message formatting, system prompt or tool definitions that the API adds to your billed total.
How do I get the exact token count?
Send the request and read the usage field in the response. That is the number you are billed for, and it is the only one that settles arguments. Any offline estimate, including this one, is a planning figure.
Why is output more expensive than input?
Generating text requires a forward pass per token produced, while input is processed in parallel. Across the models here output costs three to eight times input, which is why controlling reply length usually saves more than shortening the prompt.
Does the language I use change the cost?
Substantially. English is the most token-efficient language for these models. The same content in Spanish or German costs roughly a third more, and in Chinese or Japanese it can cost two to three times as much.
What are cached tokens?
Parts of a prompt the provider has already processed and stored, charged at a steep discount when resent — as little as 2% of the normal rate on some providers. They are the main lever for anything with a long, stable system prompt.
Are these prices current?
They were taken from each provider's own pricing page on 2026-10-26, and that date is shown on every page here. Several carry increases that have already been announced, and those are flagged. Check the provider before committing to a budget.