Through an API, you do not pay for a model by the question. You pay by the token (a chunk of text, often part of a word), in both directions: what you send and what it writes back. Per-user subscriptions hide the same cost behind a fixed price per seat and a usage cap. A demo that costs a few cents can turn into a budget line of several thousand dollars a month once it is in production.
What a request costs
The price of a request is simple arithmetic: the tokens you send times the input price, plus the tokens the model produces times the output price. Output tokens cost more, and they also take time. The model writes them one at a time, so a long answer is a slow answer.
Key concept
Every token in and every token out costs money and time. The cost of an AI feature is the cost of one request times the number of requests. Do the math before you build, not after the first invoice.
An example from October 2026, with Claude Sonnet 5.5, Anthropic's mid-range model: 2 dollars per million input tokens and 10 dollars per million output tokens. A typical request that sends 3,000 tokens and gets 500 back costs about 0.011 dollars. At 10,000 requests a day, that is 110 dollars a day, or roughly 3,300 dollars a month.
One request costs about $0.011; 10,000 requests a day cost about $3,300 a month.
3,000 input tokens light up at once, in one computation; then a 500-token answer is written one token at a time, in 500 computations in a row.
What it means for you
For founders
Compare the cost per use with what your customer pays. If a free user fires off fifty requests a day, that user sets your margin, not your paying customer. Put quotas in place from launch.
For business leaders
In a large company, an internal assistant used by thousands of employees becomes a spending line in its own right. Ask for a cost per use and per team, on top of the overall bill. Decide, too, who has the final say when a team spends more than planned. A per-seat subscription does not settle the question: as of October 2026, Claude's Enterprise plan is billed per seat, plus usage at API rates.
For tech leads
Log the input and output tokens of every call, per feature. Set an output limit (max_tokens) and spending alerts. In a startup, one dashboard is enough to begin with. In a large company, feed this data into your existing cloud cost tracking, and check what data is sent to the vendor.
By the tenth message, you are paying again for the nine before it: 55 messages paid for a 10-message conversation.
Go deeperWhy a conversation gets more and more expensive
The model has no memory between requests. To carry on a conversation, the application sends the whole history again with every message. By the tenth message, you are paying again for the nine before it, so the cost of a conversation grows faster than its length. That is one reason to manage the context window as a budget.
The classic mistake
fAIling RULE XI
The trap: Measure ROI after 48 hours
Judging an AI feature on its first 48 hours means judging a cost and a level of use that do not exist yet. Real volume, heavy users and long conversations arrive later. Measure the cost per use over several weeks of real traffic before you draw conclusions.
Read rule XI in the manifestoTry it
Pick an AI feature that is already live, or a prototype. In the vendor's console, find the average number of tokens sent and received per request. Multiply by the published prices, then by the monthly volume you expect. Compare the result with the cost of doing the same task by hand.
Around this topic
Go deeperWhat prompt caching does to the bill
In most requests, a large part of the text sent is the same every time: the instructions, the tools, the reference documents. Prompt caching saves you paying full price to read it again. An example from October 2026 at Anthropic: writing to the cache costs 1.25 times the input price (for a five-minute cache), then each cache read costs 5% of the input price on Sonnet 5.5. Once cached, a fixed opening of 20,000 tokens drops from about 0.04 dollars to 0.002 dollars per request. At 10,000 requests a day, that saves more than 350 dollars a day.
Go deeperBatch processing
Not every task needs an answer straight away: sorting yesterday's tickets, enriching a catalog, re-running a set of tests. Some vendors offer batch processing, which is cheaper in exchange for a delay. In October 2026, Anthropic's batch API is billed at half price, for input and output alike. According to its documentation, most batches finish in under an hour, with a maximum of 24 hours.
Go deeperRouting between model tiers
A single vendor sells several sizes of model, at very different prices. In October 2026, at Anthropic, a million input tokens costs 0.10 dollars with Haiku 5.5 (for a request under 100,000 tokens), 2 dollars with Sonnet 5.5, 4 dollars with Opus 5.5 and 10 dollars with Fable 5.1. Routing simple tasks to a small model cuts the bill sharply. You still have to check, on your real cases, that the small model is good enough. Reasoning effort gives you a similar lever within a single model.
Go deeperInput and output: two prices
At Anthropic, in October 2026, an output token costs five times as much as an input token, on every tier.
Vendors do not explain this gap, but the technical reason usually given lies in how the model works. It reads all the input tokens at the same time, in a single computation. It writes its answer one token at a time, though: each token waits for the one before and needs a new full computation. A 500-token answer therefore takes 500 computations in a row. So each token written ties up the machine far longer than a token read. That is most likely why output is priced higher.
Summarizing a long document therefore costs mostly input. Writing a long text costs mostly output. Asking for a short answer or a fixed format cuts both the bill and the wait. The thinking tokens of a reasoning model are billed as output tokens too.
What to do with it
Log input and output tokens per feature, and set an output limit and spending alerts.
Compare the cost per use with what your customer pays, and set quotas from launch: a free user sending fifty requests a day sets your margin.
Manage the context window as a budget: a conversation resends its whole history with every message, so its cost grows faster than its length.