Comperence is a free price comparison for GPU rental and model APIs. The name joins the two things it compares: compute and inference. There are no ads, no affiliate or referral links and no sign-up. Links to providers are plain links; we earn nothing when you click them.
Where the prices come from
Prices are collected automatically from providers' public pricing pages and price APIs, and from open price catalogs that are themselves compiled from providers' pricing APIs. Most are checked every 6 hours; marketplaces are checked hourly, and committed-term and model API prices once a day. Every price shows when it was last confirmed: when a catalog was built earlier than our check, the catalog's own time is shown.
Marketplace prices, where many independent hosts offer the same GPU, are shown at the median of the offers available at the time, so a single unusually cheap host does not set the headline. Vast.ai's prices come from its public pricing feed, with thanks.
How prices are normalized
- Compute is shown in US dollars per GPU per hour: the instance price divided by the number of GPUs in it.
- Inference is shown in US dollars per million tokens, for input, output and cached input.
- Blended price is (3 × input + 1 × output) ÷ 4, a typical mix of input and output tokens. Use it to rank providers for one model, not to compare different models: models count tokens differently.
- Other currencies are converted to US dollars. The original price and currency show when you hover over the marker.
What the labels mean
- Last checked
- When our collector last read the price from the provider.
- Stale
- A GPU price not checked for over 24 hours, or a model price not checked for over 72 hours. Stale prices are greyed out; prices not checked for over 7 days are hidden by default.
- Spot
- Spare capacity sold at a discount. The provider can take it back at short notice.
- Committed
- A lower rate in exchange for reserving capacity for a fixed term, from one month to three years.
- Marketplace
- A listing from an individual host on a marketplace. Availability can change by the minute.
- Quantized
- The provider serves the model at lower precision than its original release (for example FP8 or INT4), which can be cheaper and slightly less accurate.
Prices are indicative
Providers change prices, regions and stock often. Always confirm the price on the provider's own site before you buy. If you spot a wrong price, tell us.