How does AI model pricing work? Input, output and cached tokens explained

Every AI pricing page looks the same: a table of model names and small dollar amounts with “per 1M tokens” printed somewhere near them. The amounts look tiny and turn out not to be, because the unit is doing the heavy lifting. This guide explains that unit and the rate card built on it, using the numbers this site reads from the vendors’ own pages and keeps, dated, in its pricing ledger.

The meter runs on tokens

A token is the chunk a model actually reads and writes: a short word, part of a longer word, a punctuation mark. In ordinary English a token averages out near three quarters of a word, so a thousand words of text is roughly 1,300 tokens. Vendors publish rates per million tokens, which is about the length of a long novel, and bill you for the exact count each request uses.

The unit favours honesty in one way: it measures work, not messages. A one line question costs almost nothing; pasting a contract in costs what the contract weighs. The same logic runs the subscription products, where the plan limits that govern ChatGPT, Claude and Gemini are the vendor rationing this same underlying meter, in ranges rather than dollars.

Input, output and cached, the price lines on every page

Each model on a vendor’s page carries a small rate card, and the lines mean different things.

Input is everything you send: your question, the document you pasted, the conversation so far. This is the headline rate. On the pages we read on 6 September 2026, OpenAI prices GPT-6 Astra input at $10 per million tokens, Anthropic prices Claude Sonnet 5 input at $2, and Google prices Gemini 3.8 Flash input at $0.75. Read the conditions attached to a headline rate before you plan around it. Google’s page prints that $0.75 as running through 31 December 2026, with $1.50 starting on 1 January 2027, so the cheapest figure in this paragraph is a promotion with an expiry date on it. OpenAI’s page has a condition of a different kind: a second set of columns for long context, where the GPT-6 Astra input line reads $20 rather than $10.

Output is everything the model writes back, and it is always dearer. The same pages price GPT-6 Astra output at $50 per million tokens against its $10 input, Claude Sonnet 5 at $10 against $2, and Gemini 3.8 Flash at $3.75 against $0.75, the last of those becoming $7.50 against $1.50 when the promotion ends. Each of those flagship pairs is exactly five to one, which is a tidier picture than the market gives. Across the priced models in our ledger the multiple runs from roughly level at the cheap end to more than eight times at the steep end, and it clusters between three and five. OpenAI’s own page sits at the steep end of that spread: several of its rows are six times or more.

Cached input is the quiet discount. When your request begins with material the vendor processed moments ago, the same long document, the same instructions, it can skip the repeated work and charges accordingly. OpenAI lists cached GPT-6 Astra input at $1 per million, a tenth of fresh. Anthropic goes further: its page lists cache reads for Claude Fable 5.1 at $0.25 per million against $10 fresh, a fortieth. For any tool that keeps sending the same context, which is most assistants and nearly every agent, this line decides the real bill.

One caution from our ledger: the prices are per model, not per vendor, and the range inside one vendor dwarfs the range between them. Among the OpenAI models this site tracks, input runs from $0.20 per million tokens to $10, a factor of fifty. The vendor’s own page is wider still in both directions, from $0.05 per million on its smallest current model to $150 on the dearest reasoning model it still lists. Those are two different claims, and it is worth knowing which one you are reading: the ledger’s range covers the models a buyer is realistically choosing between, while the page’s range includes the small print at either end.

Why output costs more

Reading and writing are different amounts of work. A model digests your input in one pass, but it writes its answer token by token, each one requiring a full trip through the model with everything that came before. Longer answers also occupy the vendor’s hardware for longer. The published ratios put a price on that asymmetry, and it points to a practical rule: what you ask for shapes your bill more than what you send. Requesting a one paragraph summary instead of a rewrite of the whole document cuts the expensive line, not the cheap one.

What it means for you

If you pay a monthly subscription rather than API rates, this meter still governs your experience. The allowances on paid plans, compared on our plans page, are denominated in messages precisely because each message’s token cost varies, which is why a vendor publishes a range rather than a number. OpenAI’s Codex pricing page prints a table of local messages per five hour period, and the GPT-6 Astra row on the Plus plan reads 5 to 45. Read that label as carefully as the figure: it describes Codex usage drawn from the ChatGPT plan allowance rather than a general cap on chat, and our plans ledger keeps the qualification beside the row for exactly that reason. Long documents and long conversations spend your allowance faster because they are more tokens, not because the vendor is being arbitrary.

If you are choosing an API vendor, compare like with like. Match the input line against the input line, check the output multiple, and read the caching terms if your use repeats context, because a generous cache rate can matter more than a cheap headline. And check the date on any figure you are shown: these pages move without announcements, which is why every number in our ledger carries the day we read it.