AI API Token Costs Explained

Your AI app worked great in testing, so why did the bill jump at scale? The answer is tokens. When you call an AI API, you do not pay per question or per answer. You pay for the tokens that go in and the tokens that come out. Once you see what a token is and how the price per 1M tokens works, your AI API token costs stop feeling like a mystery and start looking like simple math you can plan around.

What a Token Is

A token is the basic unit an AI model reads and writes. It is not a whole word and not a single letter. It is a chunk of text that sits somewhere in between.

As a rough rule, one token is about 4 characters of English, or roughly 0.75 words. So 1,000 tokens is about 750 words, and a short paragraph might be 100 tokens. Common words often map to one token, while long or rare words can split into several.

This matters because the model counts tokens, not words. Both your prompt and the model’s reply are measured in tokens, and that token count is what you are billed on.

Tokens also cover more than plain words. Spaces, punctuation, code, and numbers all take up tokens too. That is why a token estimate is a close guide, not an exact character count.

How Pricing Works

AI providers quote a price per 1M tokens, and they split it into two rates. One rate is for input tokens, the other is for output tokens.

Input tokens are everything you send: your question, your instructions, any examples, and any background text you paste in. Output tokens are what the model writes back to you. The two are metered separately and added together on your bill.

Here is the part that surprises people: output usually costs more than input, often several times more. The model works harder to generate text than to read it, so each word it produces is pricier than each word you send. That is why long, chatty replies can cost more than long prompts.

The exact numbers differ by model and provider, and they change often. A smaller, faster model costs far less per 1M tokens than a large flagship model.

Quoting a price per 1M tokens keeps the numbers readable, since a single token costs a tiny fraction of a cent. You still pay only for the tokens you actually use.

A Worked Cost Example

Let us price a single request with made-up but realistic rates. Say input costs $3 per 1M tokens and output costs $15 per 1M tokens. These are illustrative examples only, not any real price.

Imagine one request sends 1,000 input tokens and gets back 500 output tokens. You price each side separately, then add them.

Input: (1,000 / 1,000,000) x 3 = $0.003. Output: (500 / 1,000,000) x 15 = $0.0075. Total: 0.003 + 0.0075 = $0.0105, or about one cent per request.

Notice the 500-token reply cost more than the 1,000-token prompt. That is the higher output rate at work. Always check the current pricing page, because real rates vary by model and provider and can change at any time.

You can run the same math in reverse to set a budget. Pick a monthly dollar limit, then divide by your cost per request to see how many calls that buys. This turns a vague worry into a clear number.

Why Costs Add Up at Scale

One cent per request sounds tiny, and on its own it is. The bill grows because real apps make many requests, not one.

Take that same one-cent request and run it 100,000 times: 0.0105 x 100,000 = $1,050. The per-call price never changed. The volume did. This is the gap between a cheap test and a busy production app.

Three things push the total up the fastest: the number of requests, the length of the context you send each time, and the length of the output you let the model produce. Long chat histories are a common culprit, because you resend the whole conversation as input on every new turn. This scaling effect is similar to how crypto gas fees explained stack up when network activity is heavy.

How to Estimate and Control Cost

Estimating is the same math every time. Count the input and output tokens for a typical request, price each by its rate, add them, then multiply by your expected request volume.

Once you have an estimate, a few habits keep the bill in check:

• Shorten your prompts. Send only the context the model truly needs, not whole documents.
• Cap the output. Set a maximum output length so replies cannot run long and expensive.
• Cache repeated context. If you send the same instructions every call, caching can cut what you pay to resend them.
• Pick the right model. Use a smaller, cheaper model for simple tasks and save the flagship for hard ones.
• Batch your requests. Grouping jobs can lower overhead and sometimes unlocks cheaper batch rates.

One more habit helps most: measure before you scale. Log the real token counts from a few typical calls, then project from those numbers rather than a guess. To turn these habits into real numbers for your own case, try our AI and crypto calculators and plug in your token counts and rates.

Frequently Asked Questions About AI API Token Costs

Sources

This article is for general education only, not financial advice. AI provider prices change often, so always check the current pricing page before you rely on any estimate. Reviewed for accuracy by Prof. Dr. Khalil Mudassar, PhD. Last updated September 12, 2026.