How Much Does the AI API Cost? A 2026 Breakdown
Per-token pricing sounds simple until the bill arrives. Here is how it works, and how to see the cost coming.
AI APIs bill you per token, the small chunks of text a model reads and writes. A single call costs anywhere from a fraction of a cent to a few cents, depending on which model you pick and how much text goes in and comes out. There is no flat fee and no per-request charge. You pay for the words, and the meter runs on both what you send and what the model says back.
The short answer
You are billed by the token. Every provider, OpenAI, Anthropic, Google and the rest, quotes a price per million tokens, and your bill is simply the tokens you used multiplied by that rate. For a small, fast model a typical question-and-answer costs a fraction of a cent. For a large reasoning model doing a long task, one call can reach a few cents. Nothing is free, but nothing is expensive on its own either. The cost creeps up through volume: thousands of calls a day, long prompts, or answers that ramble.
The rest of this post unpacks how that per-token price works, why the two halves of a call are priced differently, roughly what the tiers cost in 2026, and how to see a bill coming before you build.
How per-token pricing works
A model does not read words. It reads tokens, which are short chunks of text. One token is about four characters of English, so it works out to roughly three-quarters of a word. The sentence you just read is around a dozen tokens. If that split is new to you, the token counter shows any text broken into its actual tokens.
Here is the part that trips people up: input and output are billed separately, at different rates. Input is everything you send, your prompt, the system instructions, any files, and the conversation so far. Output is what the model writes back. A provider might charge, say, a couple of dollars per million input tokens and ten dollars per million output tokens. Same model, same call, two different meters. Your total for a call is the input tokens at the input rate plus the output tokens at the output rate.
Why output costs more than input
Output is priced higher because it is harder to produce. When a model reads your prompt, it can process the whole thing in a single pass. When it writes a reply, it generates one token at a time, and each new token needs another full run through the model that accounts for everything written so far. That step-by-step generation is slower and more compute-hungry, so it costs more.
The gap is fairly consistent. Across the big providers in 2026, output runs about four to five times the price of input. The practical lesson: a long prompt is cheaper than you might fear, but a long answer is where the money goes. If you can ask the model to be brief, you cut the expensive half of the bill.
Rough price tiers in 2026
Prices shift constantly, so treat these as a rough map rather than a quote. Broadly, models fall into two camps.
The cheap tier is the small, fast models, the ones with names like “mini” or “flash.” In 2026 these sit around fifteen cents to a dollar per million input tokens, with output a few times higher. They are built for high volume and simple tasks: classifying text, extracting fields, short replies. For most everyday work they are more than good enough.
The premium tier is the large and reasoning-focused models. These run roughly two to ten dollars per million input tokens, and output can reach twenty-five to fifty dollars per million. You pay for stronger reasoning on hard problems, longer chains of thought, and better reliability on tricky tasks. The difference between the tiers can be ten to fifty times for the same words, so picking the right model matters more than almost anything else. To compare current numbers side by side, use the model comparison, and for a fuller look at how the big three stack up, the model comparison post walks through where each one wins.
A worked example
Numbers make this concrete. Say you run a support chatbot that handles 1,000 conversations a day. A typical exchange sends about 500 input tokens (the user question plus a short system prompt and a little history) and gets back about 200 output tokens.
| Item | Cheap model | Premium model |
|---|---|---|
| Input rate (per 1M) | ~$0.15 | ~$3.00 |
| Output rate (per 1M) | ~$0.60 | ~$15.00 |
| Cost per conversation | ~$0.0002 | ~$0.0045 |
| 1,000 chats a day | ~$0.20 | ~$4.50 |
| Roughly per month | ~$6 | ~$135 |
The maths is not complicated. On the cheap model, 500 input tokens at fifteen cents per million is about $0.000075, and 200 output tokens at sixty cents per million is about $0.00012, so a single chat lands near two hundredths of a cent. Multiply by a thousand chats and thirty days and you get a few dollars a month. The same traffic on a premium model costs more than twenty times as much. Neither is scary, but the model choice clearly drives the total. Plug your own token counts and volumes into the LLM cost calculator to see your real numbers.
What drives your bill up
If a bill comes in higher than expected, it is almost always one of three things.
Long context is the usual culprit. Because chat APIs are stateless, you resend the whole conversation on every turn to keep the model in the loop. A thread that started at 200 tokens can be 5,000 tokens twenty messages in, and you pay for all of it each time. The same goes for stuffing a big document into every call. If you are unsure how much you are sending, the context window guide explains what actually goes into a request.
Big system prompts are the quiet version of the same problem. A detailed instruction block might be 1,000 tokens, and if you prepend it to every single call, you are paying for those tokens over and over. Verbose output is the third. Ask a model to “explain in detail” and it happily writes three paragraphs where one would do, on the pricier output meter.
Ways to cut the cost
A few habits keep the bill low without hurting quality. Start with the model. Send easy tasks to a cheap model and save the premium one for the genuinely hard requests. This single choice usually saves more than all the others combined.
- Trim your prompts. Cut filler from system prompts and drop old messages from long chats. Every token you remove is a token you stop paying for on every call.
- Cap the output. Set a maximum length so the model cannot run on, and ask directly for a short answer. Since output is the expensive half, this bites.
- Use prompt caching. If you send the same long system prompt every time, caching lets the provider bill that repeated part at a steep discount, often well over half off.
- Batch what can wait. Many providers offer a batch mode at around half price for work that does not need an instant reply, like overnight processing.
None of this requires clever engineering. Pick the right model, send fewer tokens, and keep answers tight, and the API stays cheap even at real volume. When you are ready to size it for your own use, the LLM cost calculator turns your traffic into a monthly number in a few seconds.
Frequently asked questions
How much does the ChatGPT API cost?
It depends on which model you call. In 2026 a small model like GPT-4o mini runs on the order of cents per million tokens, while a full GPT-4o call is a few dollars per million input tokens and more for output. A short question and answer is usually a fraction of a cent. Prices move often, so check OpenAI's pricing page for the live figures.
Why are output tokens more expensive?
Output is generated one token at a time, and each new token requires a fresh pass through the model. Input is read in one go, so it is cheaper to process. Across the major providers, output is priced roughly four to five times higher than input, which is why a chatty model can cost more than a wordy prompt.
Is the API cheaper than a subscription?
It depends on how much you use it. A flat subscription like ChatGPT Plus is a fixed monthly fee for one person clicking around a chat window. The API bills per token and suits apps, automation, or heavy batch work. For light personal use the subscription is usually cheaper. For a product serving many users, the API almost always wins on control and cost.
How do I estimate my cost?
Take your typical prompt and answer, count the tokens in each, multiply by the per-token input and output prices, then multiply by how many calls you expect. The LLM cost calculator on this site does the arithmetic for you, and the token counter tells you how many tokens a sample prompt actually uses.
What is prompt caching?
Prompt caching lets a provider reuse the processing it already did on a repeated chunk of text, like a long system prompt you send on every call. The cached part is billed at a steep discount, often 50 to 90 percent off the normal input rate. It only helps when the same text really does repeat across requests.
Do longer conversations cost more?
Yes. Most chat APIs are stateless, so to keep a conversation going you resend the earlier messages every time. That history counts as input tokens on every new call, so a long thread quietly gets more expensive with each turn. Trimming or summarising old messages keeps the bill in check.
The Scroll Team writes the lessons inside Scroll: Learn AI, a microlearning app that teaches how AI works in one minute a day. We read the papers and release notes so you do not have to.