✓ Last verified: 2026-07-27· Model: accuracy-adjusted effective cost· report an issue →
Token price is the cost of a call, not a usable result. A cheaper model that is right less often makes you pay for retries — so its true cost per correct answer can beat a pricier model, or lose to it. Enter each model's cost per task and accuracy to see the accuracy-adjusted cost, the break-even accuracy, and which model is actually cheaper per correct output.
—
cheaper per correct
—premium / correct
—cheaper / correct
—break-even accuracy
Sticker price vs the price of a result
Each model priced two ways: the raw cost per call, and the effective cost once you divide by accuracy to pay for the retries. The "retries" column is one over the accuracy — how many attempts it takes on average to land one good answer. The cheapest-per-correct row is highlighted; it is not always the cheapest sticker.
Model
$/call
Accuracy
Attempts/correct
$/correct
Monthly
How much accuracy the cheaper model needs
The cheaper model's effective cost at a range of accuracies, against the premium model's fixed cost per correct. The row where they cross is the break-even accuracy — above it the cheaper model wins, below it its retries make it the pricier choice.
Cheaper accuracy
Cheaper $/correct
Premium $/correct
Winner
The discount that costs you money
Every model-selection dashboard sorts by price per million tokens, and every one of them is answering the wrong question. The cost that lands on your bill is not the cost of a call, it is the cost of a call you can actually use — and the moment your task has a right and a wrong answer, those two numbers come apart. A model that is 30% cheaper per call but ten points less accurate is not 30% cheaper; once you retry or throw away its mistakes, you are buying more attempts per usable result, and the arithmetic can flip the ranking entirely. The fix is one division: take the cost of an attempt and divide it by the probability the attempt is good, and compare models on that. The break-even this calculator prints is the whole decision in a single number — the accuracy a cheaper model has to clear before its discount is real rather than a tax you pay in retries. This assumes you can tell right from wrong and simply re-run the failures, which is exactly the situation when you have a test suite, a verifier, or human review; if bad answers ship undetected the cheap model is worse still, because now you also pay the cost of being wrong downstream. Price a single call first with the token cost calculator, compare raw model prices on the model comparison page, and if you are chaining calls into an agent, size the full-task token burn with the AI agent cost calculator.
It takes each model's cost per task and its accuracy as a success rate. The effective cost per correct answer is the cost per task divided by the accuracy fraction — because on average you run one over the accuracy attempts to land a good answer, and you pay for the failed attempts too. Monthly effective spend is that per-correct cost times the number of correct answers you need per month (correct answers per day times 30.4). The break-even accuracy is the pricier model's accuracy scaled by the ratio of the cheaper model's cost to the premium model's cost — the accuracy at which the cheaper model's retries exactly cancel its price advantage. This assumes independent retries and a reliable way to detect a wrong answer; without a verifier the effective cost is a lower bound.
Frequently asked questions
What is cost per correct answer and why does it matter more than token price?
It is what you pay to get one usable result, not what one call costs. Token price says nothing about whether the attempt was right. If a model succeeds 70% of the time and you retry the failures, you pay for about 1.4 attempts per good answer, so its effective cost is the sticker divided by the accuracy. That division can flip the ranking: a model 40% cheaper per call but less accurate can be more expensive per correct answer than a pricier, more reliable one.
How do I calculate the accuracy-adjusted cost of a model?
Divide the cost of one call by the success rate as a fraction. Two cents at 95% is about 2.1 cents per correct answer; two cents at 65% is about 3.1 cents, because ~1.5 calls are needed per good one. The multiplier is one over accuracy — 95% is a 1.05x tax, 65% a 1.54x tax, 50% doubles your cost. It assumes you can detect wrong answers and retry; if bad answers ship, the real cost is higher.
How much accuracy does a cheaper model need to actually be cheaper?
At least the pricier model's accuracy scaled by the price ratio. If the premium model is 95% accurate and the cheaper one costs 70% as much, the cheaper needs ~66.5% accuracy to break even. Below that its retries eat the saving; above it, it wins. A headline discount is meaningless without an accuracy number attached.
Does this apply to classification, extraction and agents too?
Anywhere you can score whether an output is right. Accuracy is your pass rate against a labelled set, or the fraction of agent runs that reach a correct result. For agents the effect is amplified — a failed run can burn far more tokens than one call. For open-ended generation, use the acceptance rate (share kept without a rewrite) as the accuracy input.