xpu live / INFORMATION
How we collect & compare
Published prices, organized with the context you need to compare them.
1. Start with published sources
We collect selected information from provider pricing pages, public catalogs, documentation and APIs. Some records are gathered automatically; others are curated from published sources. Source links and checking dates are shown where available.
The catalog is a selection. It does not necessarily cover every provider, model, region, discount or deployment option. A provider being listed does not mean all of its prices have been collected.
2. Keep the billing basis visible
We organize rates into explicit units such as USD per million input tokens, per million output tokens, per image or per GPU-hour. When a rate is converted to another time unit, the conversion describes the listed usage rate, not a complete invoice.
- Input, output and cache charges are separate components.
- Standard, batch, priority and time-dependent tariffs can have different conditions.
- GPU rates can exclude CPU, memory, storage, networking and taxes.
- Image, video and audio costs can depend on resolution, duration and other generation settings.
“Lowest input price” is not “lowest total cost.”
The lowest-input token tariff is selected by its input price, with the output price from that same tariff. A different input/output mix or a conditional discount may change which option costs less for your workload.
3. Show dates and uncertainty
A checking or observation date describes when the information was captured. It is not a promise that the provider still offers the same rate or has capacity available. Automated updates depend on the collection service and source availability; they are not a continuous live quote.
If a collection attempt fails, a previous rate may remain visible with its recorded date and status. Automatically collected prices older than the current freshness threshold of 48 hours are marked stale. Manually curated listings carry their own checking dates. Unknown prices are left unpriced; an explicit zero represents a published zero rate.
4. Compare like with like
Use the provider, model version, deployment configuration, billing unit and tariff conditions together when comparing listings. “From” prices are starting rates. Account-specific discounts, negotiated contracts, credits and final taxes may differ from published rates.
GPU hourly equivalents describe the stated configuration and rental type. On-demand, serverless, spot and reserved offers are not interchangeable. A listed hardware option does not guarantee inventory.
Benchmarks and specifications
Where benchmark results are shown, they are dated results from the named source and evaluation setup. They are not xpu live test results or a universal measure of model quality. A price may refer to settings different from those used in a benchmark. Missing scores do not mean a model is weaker.
Check the original before committing
Listings are informational comparisons, not quotations or recommendations for a particular purchase. Check the provider’s official page for current prices, eligibility, service terms and availability. If you spot a mismatch, send a catalog correction with the page and source.