AI cost calculator for prompts and agents

How to use the AI cost calculator

1. Enter the number of people, the requests each sends an agent on a working day, and the working days in a month.
2. Pick the model: Claude Haiku 5.5 or Claude Sonnet 5.5.
3. Move the slider to the share of vague requests in your team: "improve it", "make it better", requests that leave the agent to guess.
4. Read the month, the year, and what the vague requests add compared with every request precise.

How the AI cost calculator works (no AI model is called)

The calculator multiplies measured figures; it runs in your browser.

• Tokens per request, measured on our platform with a real agent, in the same session: the precise request took about 49,000 tokens, "improve it" about 113,000 (four actions: the agent restyled the page). A request costs that much because an agent re-reads its instructions at every step it takes: about 20,700 tokens per step on the agent of our training sessions, so a single step costs at least about 21,000 tokens.
• Prices: the models' public list prices per million tokens (October 2026). Claude Haiku 5.5: $0.10 input, $0.125 written to the prompt cache, $0.01 read from it, $0.50 output. Claude Sonnet 5.5: $2, $2.50, $0.10 and $10.
• How a request's tokens divide: most of what an agent re-reads at each step comes from the prompt cache, and it writes little (a typical step reads about 21,000 tokens and writes about 100). The calculator prices 96% of the tokens as cache reads, 3.5% as cache writes and 0.5% as output. This puts a Sonnet 5.5 token at about 14 times a Haiku 5.5 token, in line with the 12 to 14 times we measured on a typical step.

It is an estimate: real requests vary with the task, the files the agent reads and how many steps it takes.

What the result means

The gap between the two bars is what vague requests cost on top. It grows with the team: the same habit, multiplied by every person and every working day. Asking precisely the first time (the context, the task, the guidelines, the constraints) saves more than any choice of model can.

The model matters too. For a small change, a bigger model often gives the same result for many times the price: on our measurements, the same small change cost Sonnet 5.5 about 12 to 14 times what it cost Haiku 5.5.

To check your own requests, try the prompt checker; to see what a text weighs, the token counter.

FAQ

Is the AI cost calculator free?

Yes, with no account. It runs entirely in your browser.

Where do the figures come from?

From requests measured on Tokenacademy's platform, where every request to the agent is counted step by step, and from the models' public list prices. No figure is fetched or computed by an AI model when you use the calculator.

Why does one request cost tens of thousands of tokens?

An AI agent works in steps (reading a file, making a change, checking the result) and at every step it reads its instructions again, about 20,700 tokens on the agent of our sessions, plus the conversation so far. A request of a few words can still take two to five steps.

Why are vague requests so much more expensive?

A vague request leaves the agent to guess: it reads more, changes more, and often needs another message to correct it. In our measurements, "improve it" took 113,000 tokens where the precise request took 49,000.

Are the prices in dollars?

Yes: the models' list prices are published in US dollars. Your contract or provider may differ.

How do we bring the share of vague requests down?

By practice: each employee writing real requests to a real agent and seeing what each one costs. That is what Tokenacademy's training does, with every token counted. Book a demo.