Methodology
Every number on this site comes from the automated pipeline described below. Nothing is entered by hand, which is why a price that looks wrong is a pipeline bug we can fix rather than a stale article.
1. Sources
Two independent public sources are ingested on every refresh and merged per model:
A community-maintained price registry gives the widest coverage of per-token rates, context windows and capability flags. A second live API gives per-model retail prices and descriptions. Where both cover the same model, both prices are published side by side as a cross-check.
2. Normalisation
- All prices are converted to USD per 1,000,000 tokens. Rates published per token or per 1K tokens are converted at ingest.
- Hosted copies of a model (a model served by a third-party inference platform) are kept under the platform, not merged into the model owner, so first-party and platform pricing never get confused.
- Duplicate records for the same model are de-duplicated by keeping the most complete record (prices present, context known, cross-check available).
- Fine-tuned variants (
ft:entries) and internal test records are excluded.
3. Rankings and derived numbers
- Blended cost = 3 × input price + 1 × output price. Used where a workload mixes prompt and completion tokens, because most applications send far more input than output.
- Monthly estimate = requests/day × 30 × ((prompt tokens × input price) + (completion tokens × output price) + cache-adjusted prompt tokens) ÷ 1,000,000.
- Use-case shortlists apply a hard capability filter first (context size, vision, tool calling, cache pricing), then rank by the cost profile that use case actually pays.
4. Freshness
The pipeline is scheduled to run every six hours. A run that fails keeps the previous values rather than publishing a gap, and the timestamp on each page reflects the last successful refresh — currently 2026-09-13 13:40 UTC. Changes between runs are published in the changelog and pushed to search engines.
5. Limitations — read this before budgeting
- List prices only. Enterprise agreements, committed-use discounts, volume tiers and regional multipliers are not reflected.
- Multimodal inputs (images, audio, video) are often billed by a different unit that is not comparable to text tokens; only the text token rate is published here.
- Batch, priority and offline tiers can differ substantially from the interactive rate.
- Free tiers and promotional pricing are excluded from rankings so they cannot distort the picture, but a zero price is shown as “free” when a provider publishes one.
- Providers can and do change prices without notice. Always confirm on the provider pricing page before committing to a budget.
6. Reporting an error
Use the contact page with the model slug and the correct rate — including a link to the provider page — and we will fix the mapping.