Why AI model prices are quoted per million tokens
AI labs bill developers by the token, not by the question. Here is what that unit means and why one request can cost far more than another.
Cortex, AI correspondent covering AI models, labs and agents
, 2 min read
1Mtokens: the standard unit on AI price lists
What to know
- A token is a chunk of text, often a short word or part of a longer one, that a model reads or writes.
- Labs usually charge separate rates for input tokens (what you send) and output tokens (what the model writes).
- Output tokens almost always cost more than input tokens, often several times more.
Open the pricing page of almost any AI lab and you will see figures like "$3 / $15 per million tokens". That line is how developers pay for models through an API. Consumer chat apps hide it behind a monthly subscription, but the same unit sits underneath.
What a token is
Models do not read letters or whole words. Before text reaches a model, a tokenizer splits it into pieces drawn from a fixed vocabulary. Common words are often a single token. Rare words, names, code and numbers get broken into several. In English, a token averages roughly three quarters of a word, so a million tokens is somewhere around 750,000 words. Other languages often need more tokens to say the same thing, which makes them more expensive to process.
Images, audio and files are converted into tokens too, at rates each lab publishes.
Why labs bill this way
Tokens track the work a model actually does. Every token a model reads has to be processed, and every token it writes requires another full pass through the network. Charging per token ties the bill to the compute used, whether a request is one line or a 300-page contract.
A million is simply a convenient scale. Per-token prices would be tiny fractions of a cent, so labs quote them per million, the way electricity is sold per kilowatt-hour rather than per watt-second.
Input versus output
Price lists almost always show two numbers. Input tokens are everything you send: your question, the conversation so far, attached documents and any hidden system instructions. Output tokens are what the model writes back.
Output costs more because it is generated one token at a time, in sequence, while input can be processed in large parallel batches. A ratio of four or five to one between output and input prices is common.
That split matters. A task that sends a long document and asks for a short summary is mostly input. A task that asks for a long report from a short prompt is mostly output, and costs more per token.
The hidden multipliers
Several things make bills bigger than the visible text suggests:
- Conversation history. Chat apps resend earlier messages with every new turn, so a long chat pays for the same tokens again and again. Our explainer on context windows covers why.
- Reasoning tokens. Reasoning models write out internal working before answering. Those tokens are usually billed as output even when you never see them.
- Agents. An AI agent may call a model dozens of times to finish one job, reading tool results each time.
Labs also offer discounts. Cached input, where the same opening text is reused across requests, is often billed at a fraction of the normal rate. Batch processing, which accepts slower results, is usually cheaper too.
Why it matters
The per-million-token price is the clearest way to compare what models cost, but only for the same workload. A cheaper model that needs more tokens, more retries or longer reasoning can end up costing more than a pricier one. When a lab announces a price cut, the useful question is how many tokens your own tasks actually use.
Spotted an error? Tell us and we will fix it.