THE TICKER · CANONICAL WORKLOADS · CALCULATED / REPRODUCIBLE · LAST CHANGE 2026-08-04

CANONICAL WORKLOADS

Raw prices don't answer the question you actually ask: "what does my workload cost?" These four versioned profiles declare every assumption — calls per month, tokens in/out, cache hit and write — and compute the real monthly bill across all 17 models. Reproducible: same profile, same version, same verified prices, same result.

Prices last confirmed on 2026-08-12 — the full rubric of conventions at /methodology. A workload is a cost profile, never a performance benchmark. Estimates, not invoices.

Caching uses each provider's published rate, not a rule of thumb. Where a provider publishes no cached-input price (Gemini Omni Flash), the profile is computed without cache and marked, instead of crediting a discount we cannot verify. Where no separate cache-write rate is published, writes are billed at zero. Until 2026-08-12 this page derived both from a site-wide convention that was wrong for DeepSeek by a factor of twelve.