What is AI usage pricing? Explained with a taxi meter
AI services charge by usage: a meter counts the tokens a model reads and writes, at a lower rate for input and a higher one for output. Explained with a taxi meter that counts words.
- Author
Umair Mateen- Published
- Reading time
- 5 min
- Category
- AI
In one sentence: most AI services sold to businesses charge by usage, counting the tokens (small pieces of text) the model reads and writes, with a lower price for what you send in and a higher price for what it writes back.
The everyday picture
Picture a taxi with an unusual meter. It does not count kilometres. It counts words. Every word of directions you give the driver makes the meter tick. Every word of the journey the driver covers makes it tick again, at a higher rate.
A rate card on the window shows both prices. A bigger, more comfortable car has a higher rate card. A shared ride booked for later costs less. At the end, the fare is simply the clicks multiplied by the rates.
The AI service is the taxi company. The model is the driver. The words the meter counts are tokens, and the rate card is the provider's price list.
Your directions (input tokens) and the driver's journey (output tokens) both make the meter tick. The rate card sets a price for each, and together they make the fare.
How it works, step by step
- Your request (the directions you give). Your question, instructions, pasted documents and any earlier chat history are cut into tokens. These are called input tokens.
- The model works (the driver drives). The model reads the input and writes its answer, one token at a time. The answer is made of output tokens.
- The meter counts both (two counters on the dashboard). The service records how many input and output tokens each request used. Price lists, such as Anthropic's and OpenAI's, are quoted in dollars per million tokens.
- Each counter has its own rate (the rate card on the window). Output usually costs more than input. On Anthropic's current price list, output tokens cost five times as much as input tokens for each listed model.
- Extras are added (tolls on the route). Some features charge on top of tokens. Both Anthropic and OpenAI list web search as a separate charge per 1,000 searches, on top of the normal token costs.
Try it · What would you do?
The meter is running high. What should the firm do first?
A small firm's AI assistant answers customer emails. Every email is sent with a 60-page manual, the replies are long, and a monthly report runs through the same assistant. The bill keeps growing.
Pick what you would do. You can change your mind.
Where the picture stops
- No charge for waiting. A taxi meter keeps running in traffic. An AI service bills for tokens and listed extras, not for how long you take to read the answer.
- The same words can be a different number of clicks. Each model family cuts text into tokens its own way. Anthropic notes that its Claude 4.7 and later models produce about 30 percent more tokens for the same text.
- Repeated directions can be cheaper. A taxi charges full price every time. With prompt caching, text the service has stored recently is read back at a fraction of the normal input price.
Why it matters for your business
- Long answers cost more than long questions. Because output is priced higher, asking for shorter, focused replies often saves more than trimming the question.
- Waiting can halve the fare. Anthropic's Batch API gives a 50 percent discount on input and output tokens for work that does not need an instant answer.
- The car matters. Larger models carry higher rate cards. Sending simple tasks to a smaller model keeps routine work cheap.
- Chat plans and the API are different products. Anthropic says its paid Claude plans and its API are billed separately. A team subscription for staff does not cover an app your developers build.
An example
Imagine a 25-person property management firm with an AI assistant that answers tenant emails. Each email is sent with the firm's full 60-page handbook, and the assistant writes long, friendly replies. The monthly bill keeps climbing.
The office manager looks at the meter. Most of the input is the same handbook, sent again and again, so the team turns on prompt caching. They ask for replies under 150 words. Monthly summary reports move to a batch job that runs overnight. The firm still uses AI for the same work, but now each part of the fare is something they chose.
Frequently asked questions
Why do AI services charge per token?
Tokens are what the model actually reads and writes, so they track the work done on each request. Counting them lets a small request cost little and a large one cost more.
Why does output cost more than input?
Providers set it that way on their price lists. One common reason: the model can read your input in one go, but must write its answer one token at a time, which takes more computing time.
How can I estimate my AI costs?
Estimate the input and output tokens for a typical request, multiply each by its price per million tokens, then multiply by how many requests you expect. Add any extras such as web search.
Is a ChatGPT or Claude subscription billed by the token?
Chat plans for people are sold separately from the developer API. Anthropic states that API usage is billed separately from its paid Claude plans, and API prices are listed per million tokens.
Related explainers
- What are tokens in AI? Explained with word-game tiles
- What is prompt caching? Explained with a chef's prepped base sauce
- What is batch processing in AI? Explained with a print shop's drop-off tray
- Small vs large AI models, explained with a bicycle courier and a lorry
The DevDrift builds AI features with costs you can predict and track: see how.
Sources
Could you explain this to a colleague now?
Mark it and posts that mention it will show a tick. Saved in this browser only.