How APITariff collects prices
Every number on this site comes from a provider's own public pricing data, fetched daily by an automated pipeline with fail-closed quality gates.
Sources
| Source | Models | Method |
|---|---|---|
| OpenRouter | 416 | Public model API (JSON) |
| fal.ai | 264 | Public catalog API (JSON) |
| Replicate | 439 | Official model pages (published price blocks) |
| ElevenLabs | 9 | Public API pricing page |
| kie.ai | 2 | Hand-verified from public model pages |
Normalization
Prices are converted to comparable units: text models to USD per 1 million tokens (input and output tracked separately), image models per image, video per second of output, voice and audio per minute or per 1,000 characters. When a provider bills in a unit that cannot be honestly converted (credit bundles, hardware-hours), the model is skipped rather than estimated.
Quality gates
- Anchor gate: if a source's snapshot is missing its flagship model families, the whole snapshot is rejected — the source likely degraded.
- Coverage gate: a model count dropping more than 20% against the 7-day baseline blocks the write.
- Sanity gate: any single price jumping more than 80% overnight, or turning negative, is quarantined for review instead of published.
A failed gate means nothing is written for that source that day — stale-but-correct beats fresh-but-wrong.
What this is not
APITariff is not a benchmark and not a quality ranking. A cheap model can be the wrong model. Prices shown are list prices from public pages; enterprise or negotiated rates differ. Last snapshot: 2026-09-02T14:33:27Z UTC.