Price per million tokens: how AI usage costs add up
When you use an AI model through its API, you don't pay a subscription but for what you consume, expressed as a price per million tokens. There are almost always two rates: one for the text the model reads (input) and another, higher one for the text it writes (output).
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
The principle
AI APIs measure usage in tokens, chunks of text a few characters long. Since a single token costs a fraction of a cent, providers quote prices per million tokens (often written “1M” or “MTok”). Anthropic’s pricing documentation shows the most common structure: one price for input tokens, one for output tokens, and specific rates for certain options.
The calculation, step by step
The formula fits on one line:
cost = (input tokens ÷ 1,000,000) × input price + (output tokens ÷ 1,000,000) × output price
Take a fictional model priced at $2 per million input tokens and $8 per million output tokens (made-up prices for the example). A customer-support chatbot handles 1,000 conversations a day; each one sends an average of 1,500 tokens (instructions, history, question) and gets 500 back.
- Input: 1,500,000 tokens, so 1.5 × $2 = $3
- Output: 500,000 tokens, so 0.5 × $8 = $4
- Total: $7 a day, roughly $210 over 30 days
Notice that output, despite being a third of the volume, weighs more than input. To redo this calculation with real public prices, use our agent cost calculator.
What makes the bill move
- The tokenizer: the same text yields a different token count depending on the model. Anthropic, for instance, says its recent models count about 30% more tokens for identical text.
- Discounts: prompt caching makes repeated parts of a request much cheaper, and batch processing at Anthropic gets a 50% discount on both input and output.
- Reasoning: a reasoning model produces thinking tokens, billed as output.
Common mistakes and myths
- “The cheapest model per million is the cheapest to run.” A model that writes longer answers, thinks more or fails more often can end up costing more.
- “Only my question is billed.” System instructions, conversation history and attached documents are re-read, and billed, on every call.
- “The listed price is fixed.” Prices change often; check the date they were recorded.
Related terms: token, tokenizer, prompt caching, inference.
Related terms
- Token · A small chunk of text (a word, part of a word or a symbol) that an AI model reads and writes; the unit for limits and pricing.
- Tokenizer · The program that splits text into tokens (word pieces) and turns them into numbers, the only form a language model can actually read.
- API · An interface that lets one program send requests to a service, such as an AI model, and get its responses back automatically.
- Prompt caching · Reusing the start of a prompt the model has already processed, so repeated requests run faster and the reused part is billed at a lower rate.
- Inference · The moment a trained model is actually used to produce an output: every question you send to an AI triggers an inference run.
- Free vs paid AI · How free and paid AI assistant plans differ: usage limits, available models, features and privacy options.
Try it
Full glossary
Frequently asked questions
Why is output more expensive than input?
A common explanation: the model processes input in one pass, whereas it generates output one token at a time, which uses more resources per token. Providers' price lists reflect this gap, which is often large.
Is a consumer subscription billed per token?
Generally not: consumer assistant subscriptions are flat-rate, with usage limits. Per-token billing mostly applies to API access, used by developers and automation tools.
Where can I find current prices?
On each provider's official pricing page, since prices change regularly. Our agent cost calculator uses public prices refreshed every week.
Are a model's thinking tokens billed?
Yes. Providers that document it, such as Anthropic, count internal reasoning as billed output tokens. A short answer can therefore be expensive if the model did a lot of thinking.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.