The two rates on a price sheet
Input is the material sent to a model; output is what it produces. Providers generally quote US dollars per one million tokens for each. The rates can be very different. Do not multiply a combined token total by the input rate or assume an output-to-input price ratio applies to every model.
A complete text request
Suppose a request uses 2,000 input and 500 output tokens, with hypothetical rates of $3 and $15 per million. The calculation is (2,000 × 3 + 500 × 15) / 1,000,000 = $0.0135. For 10,000 such requests, budget $135. This calculation is reproducible from a usage export and published price sheet.
Find the exact SKU
A model family can include different snapshots, service tiers and pricing routes. An aggregator may offer the same model through several providers. Match the exact model ID and provider on your invoice to its price page. Record when the rate was checked; prices can change while an old blog example remains online.
Other billable categories
Cached input reads and writes may be billed at distinct rates. Batch processing may qualify for a discount. Some long-context requests use different tiers. Images, audio, web search, storage and tool calls have their own billing rules. Calculate those line items separately rather than pretending every request is just input and output text.
Forecast from observed usage
Sample real requests across short, typical and long conversations. Calculate cost for each, then weight by expected volume. Include retries and a margin for traffic growth. Use the monthly API cost estimator to test volume, and compare its result against the provider dashboard each month.