Methodology
Every figure here is traceable. This page explains where it comes from, what it has been converted into, and where it can be wrong.
Where the data comes from
The model catalogue, per-token prices, context windows, modalities and capability flags are read from the public OpenRouter API, which aggregates what each host publishes. Gateway metadata - headquarters, datacenters, privacy and status pages - comes from the same source. Positioning, category and pricing-model descriptions are written and maintained here, and every outbound link is automatically checked.
Benchmark scores, where shown, are the Artificial Analysis indices published alongside the catalogue. They are directional, not definitive.
Normalisation
- • Prices arrive as a per-token decimal and are multiplied by 1,000,000. Everything on the site is US dollars per million tokens.
- • Blended cost = input × 0.75 + output × 0.25.
- • A missing price is shown as an em dash and stored as
null, never as zero. Zero means genuinely free. - • Routing-only aliases such as
*-latestare excluded; they duplicate whichever model they currently point at. - •
:freeand:batchvariants are folded into their parent model as tiers rather than listed separately.
Known limitations
- • Prices change without notice. Always confirm at the provider before committing spend.
- • Enterprise, committed-use and volume pricing is not reflected. Published list prices only.
- • Per-request, per-image and web-search surcharges exist for some models and are not included in the blended figure.
- • Gateway price tables sample widely-hosted open-weight models rather than every model a gateway serves.
Catalogue last fetched September 14, 2026, covering 340 models across 52 providers and 117 gateways.
Frequently asked
How often do prices update?
The catalogue is re-fetched every six hours. Per-provider quotes on model pages refresh hourly. Each page states the figures it is showing at request time.
What does blended cost mean?
Blended cost weights input tokens at 75% and output tokens at 25%, a 3:1 ratio. It exists because ranking purely on input price flatters models with cheap input and expensive output, which is the most common way a headline price misleads.
Are the prices official?
They are read from the live OpenRouter catalogue, which mirrors what each host charges. Every provider page links to that provider's own pricing page so you can confirm the figure at source before committing spend.
Why do two providers charge differently for the same model?
For open-weight models, anyone can serve the weights. Hosts differ on hardware, quantization, context limits and margin, so identical output can vary several-fold in price. Model pages show every host side by side for exactly this reason.