AI Inference Cost Calculator
A model can look cheap until retrieval context, output length, billable failed attempts, and fixed platform costs meet real successful-request volume. Enter your own current rate card and observed logs to estimate the cost per successful request and a budget worth reviewing.
What successful work really costs
Use actual token telemetry, retry logs, and your current provider rate card. A low clean request price can hide expensive retry waste at volume.
- Clean token cost per request
- $0.0040
- Retry waste
- $0.50
- Monthly token cost
- $40.00
- Monthly total cost
- $240.50
- Cost per successful request
- $0.0241
- Annualized total
- $2,886.00
One successful attempt before retries and fixed costs
Estimated token spend on billable failed attempts
Successful-request token spend before retry waste and fixed costs
Successful-request token cost, retry waste, and fixed monthly cost
Total monthly cost divided by successful requests
Monthly total multiplied by 12; not a price guarantee
Copies the displayed assumptions and estimates. Answers stay in your browser.
Privacy: calculations happen entirely in your browser. The calculator does not send or save values or calculated costs. Analytics may count only a generic calculation event, without parameters.
Worked example: what do 10,000 successful AI requests cost?
Illustrative rates, not a provider quote. This example matches the calculator's starting inputs, not any particular model: 10,000 successful requests per month, 1,500 input tokens and 500 output tokens per request, $1 per million input tokens, $5 per million output tokens, and $200 in fixed monthly costs.
- Successful-request tokens: $40.00. Each clean request costs (1,500 × $1 + 500 × $5) ÷ 1,000,000 = $0.004. Multiply by 10,000 successful requests.
- Billable retry waste: $0.50. A 5% failed-attempt rate means 500 failed attempts per 10,000 successes, not 5% of all attempts. If each failed attempt bills 25% of the clean token cost, add 500 × $0.004 × 25%.
- Monthly total: $240.50. Add $40.00 in successful-request tokens, $0.50 in retry waste, and $200 in fixed costs.
- Cost per successful request: $0.02405. Divide $240.50 by 10,000 successes. The interactive result rounds this to $0.0241; the clean token price alone is not the all-in cost.
Replace the sample prices with your current rate card and the retry assumptions with billing logs: not every failed request is billed. Change the calculator inputs to compare your workload; this worked example stays fixed. With no successful requests, per-success cost is undefined even if fixed costs remain.
Turn the operating cost into a decision
Inference cost is one input to a business case, not proof that a workflow should launch. First check whether the request shape fits your provider limits, then pair the cost estimate with observed time savings, implementation cost, and a quality review.
Separate a request price from a platform budget
The calculator multiplies input tokens by the input price per million, then does the same for output tokens. It adds those two values for clean token cost per successful request, estimates retry waste from your failed-attempt rate and billable share, then adds fixed monthly cost. The final comparison unit is total monthly cost divided by successful requests.
That separation matters. A model change can reduce token cost while a monitoring minimum, vector database, or workflow platform keeps the real monthly bill from moving much.
Measure a representative week before trusting the estimate
- 1. Export actual usage. Start with provider logs, not a hand-picked demo prompt.
- 2. Split traffic by job. Retrieval-heavy, long-form, and tool-using requests often have different token shapes.
- 3. Use the right price tier. Confirm standard, batch, cached-input, long-context, and regional pricing.
- 4. Recalculate after a prompt change. More context or hidden reasoning can change output costs before users notice.
What this estimate deliberately leaves out
This is not a vendor quote. It does not automatically fetch model prices and cannot infer provider-side billing for cached-token discounts, batch pricing, multimodal tokens, web-search calls, storage, egress, taxes, committed-use discounts, or whether a failed request was billed. Use your own logs for the retry assumptions.
Check the current OpenAI pricing documentation and Gemini API pricing documentation before using any rate. Provider rate cards and billing rules change.
Cost is only one launch gate
A cheaper request is not a better workflow if it fails task quality, safety, or ownership checks. Pair the budget with an evaluation plan and a bounded pilot.