AI API Budget Calculator

What formula does the budget calculator use?

Effective requests = requests per day × (1 − cache hit rate ÷ 100). Daily input cost is effective requests × input tokens ÷ 1,000,000 × the input price, and daily output cost is built the same way with output tokens and the output price. Daily cost is their sum, monthly cost is daily × 30, and annual cost is daily × 365.

Does the cache hit rate model provider prompt caching?

No. A cache hit here removes the whole request from the bill, as a response cache that answers without calling the model would. Provider prompt caching usually still charges for cached input at a reduced price. To approximate that, leave the cache rate at 0 and enter a blended input price that reflects the discount.

Why is annual cost not exactly twelve times monthly cost?

Monthly cost multiplies the daily figure by 30 while annual cost multiplies it by 365. Twelve 30-day months make 360 days, so the annual number is about 1.4 percent higher than the monthly number × 12. Real invoices follow calendar months, which run 28 to 31 days.

Which model prices does this calculator use?

None. It has no built-in model table: you type the input and output price per million tokens yourself. Copy them from the provider’s current pricing page, because vendors change prices and retire models, and re-run the numbers when they do.

What does cost per request mean when the cache hit rate is high?

It divides the daily cost by effective requests, meaning only requests that actually reach the API. At a 100 percent hit rate there are none, so the line shows $0.000000 and the savings line shows the full uncached bill.

How should I pick average input and output tokens?

Use the mean of logged usage, not a typical or median request, because the bill follows the total. Most APIs return the token counts used in each response, so average them over a week of real traffic. Long outliers pull the mean up and are exactly what a budget needs to include.

What happens if I enter negative numbers or a cache rate above 100?

Negative values are treated as zero and the cache hit rate is clamped between 0 and 100. Blank or non-numeric fields also count as zero, so a result of $0.00 usually means an input is missing.

What is not included in the estimate?

Only token charges are modelled. Retries, failed calls that still consume tokens, image or tool-call fees, batch or volume discounts, free tiers and fine-tuned model hosting are left out, as are taxes and currency conversion.