19 sources Mis à jour 26. 8. 28. 오후 12:00 KST Variations 571 / 24h Échecs de validation 400

About

We look at the bill, not the sticker price.

"$3 per million tokens" is marketing. Your bill is decided by output tokens, and reasoning models spend several times more of them on the same job. So we run the jobs and report medians.

How we collect

  1. We read providers' own pages only. We never cite another comparison site.
  2. We collect every six hours and archive the page source as captured. You can download it from any model page.
  3. Only values passing three validation gates (schema/range, sudden-change guard, cross-source check) are published.
  4. When parsing fails we keep the last valid value. We never fabricate an estimate to fill the gap.
  5. Per-job costs come from real API calls, and reasoning tokens are counted as output.

Why Korean gets its own section

Models split Korean characters differently, so token counts can differ by a factor of two for the same sentence. Identical price sheets can therefore mean very different bills. No international comparison site measures this.

How this is funded

Advertising and voluntary support keep it free. We take no money from any provider and hold no affiliate arrangement that could influence rankings.

Contact

hello@tokendeck.app — bug reports and requests for new sources are welcome.

Emplacement publicitaire · 728×90