Count the tokens in any text, see what one call costs on each current model, then price a whole workflow by the run and by the month.
Count the tokens
Paste a prompt, a document or a reply. The count happens in your browser and nothing you paste is sent anywhere.
- Characters
- 0
- Words
- 0
- OpenAI tokens
- 0
- Claude tokens
- 0
- Gemini tokens
- 0
What it costs
One call sends your text and gets a reply back. Set the reply length and how often you run it.
| Model | Tokens in | Tokens out | Per call | Per month |
|---|
Price a workflow
A workflow is several calls in a row. Add each step, then set how many times a month the whole thing runs.
- Cost per run
- $0
- Cost per month
- $0
- Costliest step
- None
The link carries your numbers and steps. It never carries the text you pasted.
How to read your result
A token is a piece of a word. Every model reads your prompt as tokens and writes its reply as tokens, and you pay for both. The reply usually costs five times more per token than the prompt, which is why a short prompt that asks for a long answer costs more than a long prompt that asks for a short one.
The table sorts the models from cheapest to dearest for the call you described. Treat the per-call figure as the price of one attempt. If you redo the job three times before you’d send it, the real price is three times that, plus your minutes.
The per-month figure is the per-call figure times your calls. For a job you run by hand a few times a week it will be pennies on any model, and the model that needs the least editing wins. The price starts to matter when something runs thousands of times without you watching, and that is what the workflow section is for.
In a workflow, one step usually carries the cost
Most real jobs are several calls. A lead comes in, one step reads it, another looks the company up, a third drafts the reply. Each step can use a different model, and that is where the saving is.
The example loaded above is a three-step lead reply. With the numbers shown it costs 6.6¢ a run, and the research step is 60% of that, because it runs twice and reads the most text. Move that one step from Claude Sonnet 5.5 to Claude Haiku 5.5 and the run drops to about 2.8¢. Whether the cheaper model is good enough at that step is the thing to test before you switch, and I walk through that test in which AI model to use for which job.
To fill in the steps for your own workflow, paste each prompt into the counter at the top and copy the token figure down. If the workflow already runs, your provider’s usage page shows the real tokens in and out for each call. Use those. They beat any estimate. To have Claude do the sums from your own figures, the free workflow cost skill shows its arithmetic step by step.
A worked example
Take a 2,000-word client proposal. The prompt carries your notes and a past proposal, about 5,000 tokens. The reply is about 2,700 tokens. On the prices checked on 10 November 2026:
| Model | One proposal |
|---|---|
| Claude Fable 5.1 | 19¢ |
| Claude Fable 5.2 | 15¢ |
| Claude Opus 5.5 | 7.4¢ |
| Claude Sonnet 5.5 | 3.7¢ |
| Claude Haiku 5.5 | 0.19¢ |
The gap between the top and the bottom is 100 times, and it is still less than 20¢. On a proposal, pick on quality. Now take a job that tags 5,000 incoming messages a month, 1,000 tokens in and 50 out each. That is about 63¢ a month on Haiku 5.5, $12.50 on Sonnet 5.5 and $25 on Opus 5.5. Here the price is the decision, because nobody reads each result.
How the numbers are worked out
The cost is plain arithmetic: tokens in times the input price, plus tokens out times the output price, both per million tokens. The prices come from the Anthropic, OpenAI and Google price lists, and the date they were last checked is shown under the table. Two rises Google has already announced are built in: Gemini 4 Argon on 1 December 2026 and Gemini 3.8 Flash on 1 January 2027. The table switches on those dates by itself.
The token count is an estimate, and I’d rather tell you how good it is than hide that. Counting tokens exactly means either downloading a large file or sending your text to the provider. This page does neither. It counts words, letters, digits, punctuation and line breaks and turns those into tokens with weights fitted to real counts.
- OpenAI: fitted and checked against OpenAI’s own open-source tokenizer on 600 samples of prose, prompts, code, JSON, HTML and nine languages. On the 180 samples held back for checking, the typical error was 4%, and nine in ten were within 11%. The common rule of thumb, characters divided by four, was off by 17% on the same texts.
- Claude: an estimate, set 20% above the OpenAI figure. I haven’t measured it against Anthropic’s counter yet, so treat it as rough. Anthropic says its models from Opus 4.7 onward produce about 30% more tokens than its earlier ones for the same text.
- Gemini: Google’s own rule of thumb, about four characters to a token. I haven’t measured it against Google’s counter.
Three things the calculator leaves out: images and files, the extra cost of storing repeated prompt text the first time, and the hidden reasoning some models bill as output. If a model thinks before it answers, your real output tokens will be higher than your reply length suggests.
Where a token count stops being useful
The bill for the model is rarely the biggest cost of a workflow. Your time to check the output is. A model that costs ten times more and needs no editing is cheaper than one that costs pennies and needs ten minutes of your attention on every run. I cover the measuring side in how to tell if your AI agent is working, and the wider budget in how much an AI agent costs. The workflows I build get written up step by step in Bionic Business.
Common questions
How many tokens is 1,000 words?
About 1,360 for ordinary English prose, measured with OpenAI’s tokenizer on this site’s own articles. Code, tables and other languages use more tokens for the same number of characters.
Why do the three counts differ?
Each provider splits text into tokens with its own method, so the same paragraph is a different number of tokens on each. You are billed on the provider’s own count, never on someone else’s.
Is my text sent anywhere?
No. The counting runs in your browser. The share link carries your numbers and step names and never the text you pasted.
What is batch pricing?
All three providers charge half price if you send a pile of requests together and wait for the answers instead of getting them straight away. It suits jobs like tagging or summarising a backlog. It doesn’t suit anything a person is waiting on.