Usage mix shifts affect AI cost forecasts because total token volume alone does not show how costs are distributed between input and output. Measured token usage is more representative than pre-run assumptions, but it reports token counts rather than the final billed amount.
The cost estimate also depends on the two token categories and their respective prices:
approximate model cost = ((input tokens × input price) + (output tokens × output price)) / 1,000,000
How does the input–output mix change the estimate?
At the same total token count, a shift toward output tokens increases the estimate if output tokens have the higher price. It decreases the estimate if input tokens have the higher price. If both prices are equal, the mix shift does not change the approximate token cost.
| Forecast change | Effect on the approximate estimate |
|---|---|
| Input-token volume increases | Cost rises according to the applicable input price |
| Output-token volume increases | Cost rises according to the applicable output price |
| Usage shifts from input to output at an unchanged total | Cost moves according to the difference between their prices |
| Input and output volumes both increase | Both price components rise |
This means a forecast can change even when the combined number of tokens stays stable. It can also change when the input–output split remains stable but the applicable prices differ. Without both volumes and both prices, the total token count is not enough to explain the forecast.
How should operators check the forecast?
A reliable check separates volume, mix, and price:
- Start with measured token counts. Pre-run assumptions can provide a baseline, but measured usage reflects actual input and output counts more closely.
- Keep the two categories separate. Combining them may conceal a shift in usage mix.
- Apply the applicable price to each category. Input tokens and output tokens must be costed separately using the formula above.
- Compare the baseline and forecast scenarios. Review whether token growth comes from more requests, longer inputs, larger outputs, or a change in their proportions.
- Label the result as an approximation. Measured token usage is not the same figure as the final billed amount, so the estimate should not be presented as an invoice forecast.
What still needs confirmation?
The cited guidance does not establish the prices, billing rules, discounts, usage limits, or contract terms that may apply to a particular workload. Those details must be confirmed from the applicable pricing and billing materials before a forecast is treated as a spending commitment.
The central planning rule is straightforward: forecast token volume, forecast the input–output mix, and price each component separately. Any change in one of those inputs can alter the expected cost—even when total measured token usage does not move.