What Is Cost per Token in AI?

How token pricing works

When you use an AI model through an API (a connection that lets your own software send requests to the model), the provider counts the tokens in each request and response and bills you for them. (If you’re new to the unit itself, see what tokens are.) Prices are usually quoted per million tokens, because the price of a single token is a tiny fraction of a cent.

Most providers split the price in two:

• Input tokens: everything you send the model, including your instructions, any documents or data you attach, and, in a multi-turn chat, the earlier conversation, which is typically processed and billed again as input with each new message.
• Output tokens: everything the model writes back. These usually cost more per token, partly because the model can read your input all at once but has to write its answer one token at a time, which ties up the hardware longer per token.

Each model has its own rates. Larger, more capable models generally cost more per token than smaller, faster ones, so the model you pick can matter as much as how much you use it. Every request is a round of inference (the model doing its work), so your bill grows with both how often you call the model and how much text each call involves.

Pricing variations that change the bill

The basic input and output rates are only part of the picture. Many providers add variations worth checking before you estimate costs:

• Cached input discounts: if many requests start with the same long block of text, such as standing instructions or a reference document, some providers can store that repeated portion and charge a reduced rate when it is reused. Details vary: some providers charge extra to write content to the cache or to keep it stored, and cached content expires after a set time, so the discount only applies when the same text is reused often.
• Batch processing: for work that doesn’t need an instant answer, like tagging a backlog of records overnight, some providers offer a cheaper rate if you submit requests in bulk and accept results later.
• Reasoning tokens: some models work through a problem internally before answering. Those extra “thinking” tokens are typically billed as output, even when you don’t see all of them, so the same question can cost more on a reasoning model.
• Other inputs: images, audio, and very long prompts may be counted or priced differently, depending on the model.

Because these rules differ between providers and change over time, check the current pricing page for the specific model you plan to use.

How to estimate your monthly bill

A rough estimate needs three numbers: tokens per task, tasks per month, and the model’s rates. Keep input and output separate, since they’re priced differently:

monthly cost = tasks per month × (input tokens per task × input rate + output tokens per task × output rate)

If rates are quoted per million tokens, divide the token totals by one million before multiplying by the rate.

To make the estimate realistic:

• Measure, don’t guess. Run a handful of real tasks and read the token counts the provider reports for each request.
• Count the hidden input. Standing instructions and retrieved documents are easy to overlook, and a chat history grows with every turn. All of it is billed as input on each call. The more you load into the context window (the working space that holds everything the model reads for a request), the more you pay.
• Add a buffer. Retries, unusually long tasks, and growth in volume push real usage above an average-case estimate.

Cost per token tells you the raw price of the model. Whether a task is worth running with AI is a separate question about the cost of each finished task, covered under unit economics.

Subscriptions vs. usage-based billing

Not every AI bill is calculated per token. Individual and small-team plans for chat apps such as ChatGPT, Claude, and Gemini are usually flat monthly subscriptions, often priced per user, with usage limits instead of a token meter. You pay the same amount whether you use them lightly or heavily, up to those limits. Some enterprise plans pair a per-seat fee with metered usage, billed by tokens or credits.

Direct API access, which is how you or a developer connect a model to build something such as a support assistant, a document-processing workflow, or an internal tool, is typically billed by usage. There, each request is metered and charged by the token.

The two approaches suit different needs:

• Flat subscriptions give predictable costs and work well for people using AI directly in a chat interface.
• Usage-based billing scales with volume, fits automated or high-volume tasks, and lets you choose cheaper models for simple jobs. Because costs move with usage, it needs monitoring and, where available, spending limits or alerts.

Plan structures change often, so check how a specific plan charges for usage before comparing prices.

Further reading

• AI ImplementationHelpful if you’re planning to connect AI to your systems and automate work, which is where per-token costs become part of the budget.