برآورد هزینه پروژه AI

The cost of an AI project is not only model spend. Every job you give the model runs several times (error rate), someone must review the output (review), some outputs are unused (acceptance rate) and the system needs monthly maintenance — and all of this is converted to tomans at a dollar rate you choose.

This tool puts all these pieces together and shows what each “successful task” costs, what the monthly cost is, how much you save (or lose) versus the current method, and in how many months setup cost is recovered. It fetches no price or rate from elsewhere; every figure you see comes from your own inputs.

• Cost per attempt = (input tokens × price per million input tokens + output tokens × price per million output tokens) ÷ 1,000,000 × dollar rate; or a direct amount you enter (in dollars × rate, or in toman).
• Expected attempts per task = 1 ÷ (1 − error rate). At 8% error rate, each task averages 1.09 attempts; model cost per task = cost per attempt × this factor. Error rate of 100% or more is undefined and the tool warns.
• Review cost per job = review minutes ÷ 60 × reviewer hourly cost; once per job, not per attempt.
• Monthly AI cost = number of tasks × (model cost per task + review cost per task) + monthly maintenance.
• Cost per successful job = monthly AI cost ÷ (job count × acceptance rate). Rejected jobs have consumed their cost but produced no result; that is why the denominator is accepted jobs only.
• Current state: cost per job = manual minutes ÷ 60 × current hourly cost, and monthly = job count × that figure. If rejected jobs are done manually (the default), their cost is added to the proposed scenario.
• Monthly savings = current monthly cost − suggested monthly cost; setup payback (months) = setup cost ÷ monthly savings. When savings are not positive, the tool says "no payback", not a number.
• The sensitivity table recalculates all figures once with acceptance rate ±10 percentage points and once with error rate ±5 percentage points (within the allowed range).
• The dollar rate and price per million tokens are your own inputs. The dated reference table above the form is not used in any calculation.

• This tool’s output is cost, not profit: revenue, better quality, response speed, and accepted error risk are not in the number. Monthly savings compares only human cost for the same work to AI cost.
• Error rate and acceptance rate should come from a real experiment on your own data; example figures are hypothetical. A small experiment (a few hundred jobs) is better than any guess.
• Tokens per job depend on the model, prompt and language; Persian text usually uses more tokens than English. For an exact count, use the provider's own token counter.
• You choose the dollar rate and you are responsible for it; this tool neither reads nor suggests a free-market or official rate. As the rate moves, dollar costs change month to month.
• Reference prices are dated and may have changed; prompt caching, batch processing, and volume discounts are also not in this calculation. Enter the price from the official page.
• Hidden costs—data preparation, process change, team training, privacy, and compliance—belong in your setup or maintenance cost; the tool does not add them.
• This tool is a calculator, not advice: none of the figures is a recommendation to buy, contract or invest.