What an AI prediction actually costs to produce
One prediction costs us $0.01229 in model compute — about 1.2 cents, across 74 model calls per run. Even so, on a $2.99 monthly plan the card processing fee still costs about 3.1× more than a full month of AI. The most expensive part of a cheap AI product is usually not the AI.
By Zinedine · Founder & Developer · Published 26 August 2026 · Updated 2026-08-27
Correction — 27 August 2026
This page originally reported $0.000479 per prediction, and said payment processing cost 81× more than the AI. That was accurate for the engine we ran at the time: a single model call. We have since replaced it with a genuine multi-agent simulation — 24 stakeholders cast from your scenario, reacting over 3 rounds — which costs about 26 times more to run. The figures below are the new engine. We have left the old number visible rather than quietly overwriting it.
What a run actually does
A prediction is not one model call. It is 74: one call casts 24 distinct stakeholders out of your scenario, those agents then react in parallel across 3 rounds — 72 calls, with each round seeing a digest of the last so positions can move — and a final call writes the report from the transcript they produced. The whole thing takes about 29 seconds.
That structure is why cost barely tracks scenario length: the shape of a run is fixed, so what you pay for is the cast and the rounds, not how much you wrote.
Method
The measurement imports the production simulation engine directly and runs full scenarios through it, so it exercises the same code path a paying customer does rather than a reconstruction of it. Costs sum OpenAI’s own per-call token counts across every call in a run, priced at list for gpt-4o-mini: $0.15 per million input tokens and $0.6 per million output tokens.
Every run
| Run | Calls | Cost |
|---|---|---|
| medium scenario | 74 | $0.01328 |
| short scenario | 74 | $0.01159 |
| medium scenario | 74 | $0.01200 |
| Average | 74 | $0.01229 |
Range $0.01159 to $0.01328 · about 29s per run
The AI is still not the expensive part
| On one $2.99 monthly plan | Cost |
|---|---|
| Card processing — Stripe published US rate | $0.387 |
| AI compute — all 10 predictions | $0.123 |
Processing still costs about 3.1× what a whole month’s AI usage does — narrower than the 81× we reported for the old engine, but the same conclusion. The $0.30 fixed fee is charged identically on a $2.99 plan and a $29.99 one, and that is the real constraint on cheap subscriptions.
The processing figure uses Stripe’s published US rate of 2.9000000000000004% + $0.30, not our own settled fees, which vary by card country. International cards cost more, so this is a conservative floor.
What this does not include
This is the marginal cost of producing one prediction, not the cost of running the business. Excluded: hosting, database, transactional email, domains, failed and refunded payments, chargebacks, support time, and the engineering that built and maintains it. Most of those are fixed or step costs that do not fall when a customer runs fewer predictions.
We publish the compute figure because it is the number people guess hardest about and get most wrong. Having the real one in public makes every argument about AI pricing, ours included, a better argument.
Questions
How much does one AI prediction cost to run?
Measured across 3 full runs of our production engine on gpt-4o-mini, one prediction costs an average of $0.01229 — about 1.2 cents. A single run makes 74 model calls: one to cast 24 stakeholders from the scenario, 72 agent reactions across 3 rounds, and one to write the report from the resulting transcript.
What is the most expensive part of a $2.99 AI product?
Still not the AI, though the gap narrowed sharply when we replaced a single-pass engine with a real multi-agent one. On a $2.99 monthly plan, card processing costs roughly $0.39 at Stripe's published US rate, while a full month's allowance of 10 predictions costs $0.123 in compute — payment processing is about 3.1× the AI cost. The $0.30 fixed fee is charged identically on a $2.99 plan and a $29.99 one, which is what makes cheap subscriptions hard to run.
Does a longer or more detailed scenario cost more?
Less than you would expect. Cost is dominated by the fixed shape of a run — 24 agents reacting over 3 rounds — rather than by how much you wrote. Across our runs the spread was $0.01159 to $0.01328, and the most expensive run was not the longest scenario.
Why does MiroFish charge $2.99 if compute costs about a cent?
Because inference is one input among several. Payment processing, hosting, database, transactional email, domains, failed and refunded payments, support, and the engineering behind it all sit on top — and unlike compute, most do not scale down when a customer runs fewer predictions. At $0.01229 a run, compute is roughly 4.1% of Starter revenue if a subscriber uses every credit in the month. We publish the figure so that comparison can be made openly rather than guessed at.
How was this measured?
By importing the production simulation engine directly and running full scenarios through it, so the measurement exercises the same code path a paying customer does. Costs sum OpenAI's own per-call token counts across all 74 calls in a run, priced at list: $0.15 per million input tokens and $0.6 per million output tokens.
Why did this number change?
The earlier version of this page reported $0.000479, measured when a prediction was a single model call. We have since replaced that with a genuine multi-agent engine — 24 distinct stakeholders reacting over 3 rounds — which costs about 26 times more to run and produces a materially better report. The old figure is left visible above rather than quietly overwritten.
Reuse: quote or chart these figures with attribution to MiroFish and a link to this page. Model prices and engine design both change — the measurement date is stated so you can judge how stale it is, and corrections are noted on the page rather than silently applied. See also where AI prediction users come from.
See what 74 model calls actually produce — a full example report, before you pay anything.
View the example report