Know the cost of an accepted result

What an accepted result costs

This calculator answers a planning question: after retries, fixed costs and review, what does one usable result cost? Its inputs are assumptions you control. They are not a vendor quotation, a measured model benchmark or a forecast of revenue. The worked cases below were executed against this site's existing JavaScript calculation function on November 6, 2026; no paid model request was made.

Work through the default scenario

Suppose a workflow receives 10,000 original requests per month. Each uses 1,200 input tokens and 400 output tokens. For illustration, use $1 per million input tokens and $4 per million output tokens, add 10% extra requests for retries, accept 80% of the original requests, and allow $20 for fixed costs plus $100 for review. These deliberately hypothetical prices should be replaced with your own billing inputs.

Input tokens cost 10,000 × 1,200 ÷ 1,000,000 × $1 = $12.
Output tokens cost 10,000 × 400 ÷ 1,000,000 × $4 = $16.
Applying the retry factor gives ($12 + $16) × 1.10 = $30.80 in model cost.
Adding the two overhead budgets gives $30.80 + $20 + $100 = $150.80 per month.
The acceptance assumption gives 10,000 × 80% = 8,000 accepted results. Dividing gives $150.80 ÷ 8,000 = $0.01885, displayed as $0.0189 per accepted result.

Change one assumption at a time

Default assumptions: $150.80 monthly cost; 8,000 accepted results; $0.01885 per accepted result.
Acceptance falls to 40%: $150.80 monthly cost; 4,000 accepted results; $0.03770 per accepted result.
Review budget rises to $200: $250.80 monthly cost; 8,000 accepted results; $0.03135 per accepted result.
No retries: $148.00 monthly cost; 8,000 accepted results; $0.01850 per accepted result.
Zero-priced model, same overhead: $120.00 monthly cost; 8,000 accepted results; $0.01500 per accepted result.

In this example, halving acceptance doubles the unit cost without changing the bill. Removing the retry allowance saves $2.80; it does not remove the $120 overhead budget. A zero-priced model still has a nonzero unit cost when hosting and review remain. These are consequences of the stated inputs, not claims about which product performs better.

Reproduce the calculation

Download the full inputs and outputs as JSON or the seven valid scenarios as CSV. The JSON includes invalid-input checks and the calculation script's SHA-256 fingerprint. The CSV leaves the unit-cost cell empty when there are no accepted results. The exported values retain arithmetic precision; the page rounds money for display.

Try the default inputs in the calculator, then change only acceptance to 40%. Open Compare scenarios, save both sets with meaningful names and compare the results. Up to three scenarios can be stored in this browser. Export a copy before clearing browser data; saving here does not upload an account record or create a server backup.

The boundaries are part of the result

At 0% acceptance the monthly bill remains $150.80 in the default example, but the cost per accepted result is undefined. The output uses null, not zero or infinity. The executed checks also reject zero or fractional request counts, acceptance above 100%, negative costs or retries, off-step prices and nonnumeric inputs. A one-request, 100%-acceptance case with no overhead returns $0.00308; that is an arithmetic boundary check, not a realistic commercial forecast.

The retry model assumes each extra request has the same average token profile. It does not separately model cache discounts, tools, storage, failed-attempt billing or manual correction time. The review field is a monthly dollar budget, not minutes; convert your actual time and rate before entering it. Acceptance is a share of original requests, so it must already reflect your chosen quality rule after the retry process.

Turn an estimate into an experiment

Choose a fixed task and write an acceptance rule before trying a workflow. Use actual token and billing exports where available, record retries, and count the time spent checking and correcting results. Enter the observations here and keep assumptions visibly separate from measurements. Taxes, refunds and payment fees remain outside the model. The experiment-card download is a blank template; neither a saved scenario nor a passing calculation turns it into a completed customer trial.