What AI Actually Costs: The Four Numbers Behind Your API Bill

Nobody is surprised by the first AI bill. They are surprised by the third one — the month the pilot became a feature, the feature got a system prompt, and the system prompt started travelling with every single request.

The confusing part is that none of the individual numbers look alarming. A reply costs a tenth of a cent. A model charges two dollars per million tokens. A tool call is a rounding error. The bill is not made of alarming numbers; it is made of small ones multiplied by a volume nobody wrote down.

Four numbers decide it. Here is what each one does, priced on a real exchange rather than a hypothetical one.

1. The model you pick, priced on your text and not on its rate card

This is one customer-support reply: a nine-day-late order, a policy that allows a free reshipment after seven days, and an answer that apologises once, states what is being done and commits to a date. Roughly 60 tokens in, 100 out.

Two things in that table are worth more than the prices.

The first is the spread: the same reply is 17× cheaper on the cheapest model than on the most expensive one here — and this is a support message, not a research task. Most teams pick a frontier model for everything because the first prototype used one.

The second is the token column. Claude Sonnet 5 reads the same words as 73 tokens where GPT-5.6 Terra reads 56, because a token is a property of the tokenizer and not of your text. Anthropic states that Claude 4.7 and later produce roughly 30% more tokens for the same input than earlier Claude models. So a lower price per million does not automatically mean a cheaper request — you have to count the tokens the model in question would actually charge you for, which is what the AI Token Counter does across every model at once.

2. The instructions you resend 10,000 times

The reply above needs context: tone, refund policy, product notes, what the agent may and may not promise. Call it 1,500 tokens of system prompt — modest by current standards.

Nobody counts it, because it is written once. But it is sent on every request, so at 10,000 replies a month it is 15 million input tokens: more than nine times the volume of the customer messages themselves.

Caching the fixed part cuts that bill by 61% — from $44.26 to $17.32 — because providers bill a cache read at roughly a tenth of the base input rate. It is the single highest-leverage change on this list and it does not touch the model, the prompt, or the quality of the answer. You are paying full price to re-send text the provider already has.

3. Whether it is one call or a loop

An agent is where cost stops being linear, and it catches people out because a "task" still feels like one unit of work. It is not. Each step resends the conversation so far, so input grows roughly as the square of the number of steps.

Same model, same system prompt, same 10,000 tasks a month:

Twelve steps is not twelve times one step — it is seventeen times, because the twelfth call carries eleven turns of history with it. The naive estimate ("tokens per call × calls") is already wrong by half at five steps. If the loop also has tools attached — an MCP server or five — their definitions ride along on every step too, which is its own line on the bill. If you are budgeting for anything with a loop in it, price the loop: the AI Agent Cost Calculator models the growing context, the retries and the caching per call rather than per task.

4. How much the model writes

Output is charged at several times the input rate — six times, on the model above. It is also the half you control most directly, and the half people forget when they write "be thorough" into a prompt.

Tripling the reply length, with everything else identical, takes the same 10,000 replies from $44.26 to $68.20 a month — 57% more, for text nobody asked to be longer. At that length, 46% of the bill is what the model writes.

"Answer in two sentences" is a cost control. It is usually also a better support reply.

What this means in practice

• Price your own text, not a benchmark. Paste a real prompt and a real reply into the AI Cost Calculator and you will have your per-message figure in about ten seconds.
• Cache the fixed part before you switch model. A 61% cut with no quality trade-off beats most model downgrades.
• Ask what the cheap model gets wrong. The 17× spread is real, but so is the cost of a bad refund decision. Test the cheap model on your actual traffic before assuming the frontier one is required.
• Count the loop, not the task. If your feature calls the model more than once per user action, the naive number is not conservative — it is wrong.
• Then ask whether it pays for itself. Cost is only half the question; the AI ROI Calculator puts the monthly bill against the hours it takes off your team's plate, in your own currency.

If you are running AI on a consumer plan rather than an API key, the arithmetic is different again: a subscription is flat, so it gets cheaper per message the more you use it, while an API bill does not. The crossing point is a specific number of requests a month, and it moves sharply with how many people are on the plan — the subscription vs API calculator finds it for your usage.

Where these numbers come from

Every figure above was produced by the same functions that run the calculators on this site, against the published rates on the providers' own pricing pages — OpenAI and Anthropic among them. Nothing is a remembered price.

The assumptions, stated plainly so you can disagree with them:

• One support exchange of ~60 input tokens and ~100 output tokens, counted per model rather than assumed to be identical.
• A 1,500-token system prompt, resent on every request.
• 10,000 requests a month, and no batch discount.
• The agent rows use one system prompt, 300 tokens of instruction, 250 tokens written per step, and a context that grows with each step — no retries, which makes them the optimistic case.
• Token counts outside OpenAI's tokenizer are estimates with a stated range, not measurements. Only OpenAI publishes a tokenizer that runs locally; anything claiming an exact Claude or Gemini count is inventing precision.

Where the assumptions bias the result: a real deployment has retries, longer histories and more than one prompt version in flight, so the figures above are floors rather than forecasts.

Related reading

The multiply-by-volume trap is not unique to AI — it is the same arithmetic behind what your daily habits cost per year. If AI arrived on your card as a monthly plan rather than a bill, it belongs in a subscription audit with everything else that renews silently. And if you are freelancing, AI tooling is now part of what you need to charge to cover your costs.