Methodology

Sources

One rule: a rate is only used if it appears on the provider's own public pricing page. That means openai.com, anthropic.com, ai.google.dev, mistral.ai, cohere.com and the equivalent page for every other provider covered here.

We deliberately do not use:

• Rates quoted in other people's blog posts or comparison articles, which go stale silently
• Negotiated or enterprise pricing, which is not available to most readers
• Promotional or trial pricing presented as if it were the standing rate
• Anything from a leak, a screenshot or an unnamed source

How the calculators work

Every model has a separate input and output price per million tokens, and for most workloads those two differ by a factor of three to five. A calculator that blends them into one number will be wrong for any workload that is not perfectly balanced, so ours keep them apart.

The general shape of the calculation is:

1. Convert your described workload into input and output tokens per request
2. Multiply each by the provider's per-million rate for that direction
3. Multiply by request volume over the period you chose
4. Add any per-request or per-image charges the provider bills separately

Where a provider prices in a unit other than tokens — per image, per second of audio, per character — we use that unit rather than approximating it in tokens.

Token estimates

When you describe a prompt in words rather than tokens, we convert using a rough ratio of about 4 characters per token for English. That approximation is reasonable for English prose and steadily worse for code, for languages that are not English, and for text full of rare symbols.

If your cost estimate matters to a decision, count the tokens with the provider's own tokenizer rather than trusting the conversion.

How often rates are checked

AI pricing changes more often than any other software category we track. Providers cut prices, introduce cheaper tiers, and deprecate models on short notice. We re-check the pricing pages for the models used in the calculators on a rolling basis and update the figures directly on the page when they move.

Because of that pace, treat any AI pricing page — including this one — as needing verification before you commit real budget. If you find a rate here that no longer matches the provider's page, email us and it gets fixed.

What the calculators cannot know

These are the four things that most often make a real bill diverge from an estimate, and none of them can be inferred from the inputs:

• Prompt caching. Providers charge much less for cached input tokens. A system prompt re-sent thousands of times a day costs very differently with caching than without.
• Retries and failures. Timeouts, rate limits and validation failures all generate billed calls that produce nothing useful.
• Context growth. In a conversation, the whole history is re-sent with every turn. Cost per turn rises through a long session; a flat per-request estimate does not.
• Traffic shape. Averages hide peaks, and peaks are where rate limits, and the more expensive workarounds for them, appear.

As a rule of thumb, a production workload tends to run above a naive estimate rather than below it. Plan with headroom.

Corrections

Send the page and the model to shershenbizi@gmail.com. We check the claim against the provider's pricing page, correct it on the page, and update the check date. We do not quietly change a number and leave the date untouched.

Independence

No provider pays for placement, ranking or inclusion here. The site is funded by advertising, which is described on the about page, and advertising has no influence on what the calculators return.