Cost per request calculator
Per-million-token pricing is the wrong unit for a product decision. This one starts from a single call and scales it out to the numbers you actually plan against: per thousand calls, per user, per month.
How this is calculated
Cost per call is (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate). Everything below it is that figure multiplied out: per thousand calls, per user per month at your calls-per-user figure, and per month across your whole user base. No caching or batch discount is applied, so the numbers are the unoptimised baseline.
Reading the unit economics
Per-user cost is the number that decides whether a product works. It is worth calculating before you launch a feature rather than after, because the levers available afterwards are narrower than the ones available during design. Once a feature ships, you are choosing between a cheaper model and a worse product; during design you can still change how many calls the feature makes at all, which is the more powerful variable.
Watch the calls-per-user figure especially. Cost scales linearly with it, and it is the number most often underestimated: an agent loop that averages six model calls per user action costs six times what a single call does, and features like automatic retries, reranking passes, and self-critique steps multiply it further without ever appearing in a per-call estimate.
Cost per request questions
How much does one LLM API call cost?
Almost always a fraction of a cent, which is exactly why per-call cost is the number worth knowing. A typical 4K-input, 800-output call on a mid-tier model lands around a cent; on a flagship it is several cents; on a Flash-class model it is a small fraction of one. Multiply by your call volume and the differences stop being rounding errors.
How do I work out cost per user?
Set the calls-per-user field to the number of API calls an average user generates in a month. The result is your per-user monthly cost of goods, which is the figure you compare against your subscription price to see whether the unit economics work at all.
What is a healthy AI cost per user?
There is no universal answer, but the shape of the question is the same everywhere: your per-user inference cost has to leave room for everything else in your gross margin. If it is a double-digit percentage of your subscription price, the usual levers are a cheaper model tier, prompt caching, tighter output limits, and not calling the model at all when a cheaper path exists.
Why does my actual bill exceed this estimate?
Usually retries, system prompt growth, and output length. Failed calls still bill for what they generated, system prompts accumulate instructions over time, and models that are not constrained with a max output setting will happily write more than you planned for. Estimate on your real p50 and p90 token counts rather than your intended ones.
Should I price on average or worst-case token counts?
Both. Average tells you the expected bill; the p90 or p99 tells you what a bad month looks like. If those two numbers are far apart, output length is your real cost driver and capping it is more effective than switching models.