Formula. Observed blended rate = model estimated cost in USD ÷ model total observed tokens × 1,000,000. Total tokens include input, output, cache-creation (cache write), and cache-read tokens. The table shows each category as a share of the same total-token denominator.

Window and coverage. The analysis requests the 30-day public insights aggregate. Models must have a positive recorded estimate and at least one million observed tokens to enter the ranking. Pricing coverage separately shows how much of the full token cohort has a nonzero estimate; unpriced rows and sub-threshold priced rows are counted and disclosed, not assigned invented rates. Because the insights API returns a bounded model table, any token volume outside the returned rows remains outside the priced numerator and therefore lowers the coverage percentage.

Pricing freshness. Cost fields are recorded as API-equivalent estimates when usage is collected using the available public pricing table. Provider prices can change, live price matching can miss a model identifier, and a later price update does not necessarily retroactively reprice every historical row in this rolling window. Provider billing pages remain authoritative.

Interpretation. A higher blended rate may reflect a more expensive model, more output tokens, fewer discounted cache reads, or a combination of those factors. A lower rate does not establish better value or quality. Subscription products and bundled allowances can make a participant's actual marginal payment very different from the API-equivalent estimate.

Cohort. Results cover developers who choose to participate and be listed publicly. They describe this observed cohort—not the market, every user of a model, or a representative developer survey. Prompts, source code, filenames, and model outputs are not dimensions of this analysis.