The main table
- Input / Output (official): the vendor's own list price per 1 million tokens. Every other channel is compared against this reference.
- Lowest input: the lowest input price seen on any channel for that model, with the channel's name and the difference from the official price.
- Channels: how many places we hold a current price from.
- Status: Verified means an independent second reader has re-read the official price from its source; Unverified means it is recorded but not yet re-read.
Click a column header to sort. Search by model, vendor or API name, or filter by vendor and channel type.
A model page
- Official list price: all recorded rates for the vendor's own API, including cached, cache-write and batch rates. A dash means not recorded, not free.
- All channels: the latest price from each channel, its region, pricing mode, and difference from the official price. The lowest input price is flagged.
- Sources: every snapshot the page's numbers came from, with URL, fetch time, size and SHA-256.
The compute pages
- Per GPU-hour: the price of renting one GPU for one hour. A provider that sells whole machines is divided by the machine's GPU count.
- Kind: on demand, spot, reserved and marketplace prices are different products and are listed apart; the reference price uses on-demand prices only.
- Same GPU, different machine: where the provider states it, the vCPUs, memory and network that come with one GPU.
- Mainland China: prices in CNY have a section of their own, shown as published and never converted.
Comparisons and the weekly report
Each model page ends with "Compare with": two models side by side, with prices and measured speed. The blog carries a report every Monday, computed entirely from the records.
Using the data
The whole dataset is available as JSON with the same records the pages show. See Use in your AI for how to connect an AI assistant or script.
Something looks wrong?
Open the source link first: vendors change their pages without telling anyone. If the record and the source disagree, email hello@auramatrix.ai with the model, channel and link. Corrections are added as new records; the old ones stay. See Methodology.