We are rebuilding this page. An earlier version, published on 19 October 2026, showed throughput and latency figures that we have withdrawn because most of them were not measurements. This page will return only when every number on it comes from a real request that we can reproduce.
The withdrawn page showed two Groq probes of one and two requests each, with confidence intervals that so few requests cannot support. The dataset files behind it also held six more rows (Cerebras, SambaNova, AMD, Cloudflare, OpenAI) whose figures were typed into the generating script as sample values and labelled as measured. The page's claim of an 80-node measurement grid was not supported either: the run checked the health of 10 nodes and issued every request from one machine. The page, the dataset downloads and the JSON feed have been withdrawn. If you saved or cited a number from them, please do not rely on it.
Regions. Three: US West, US East and Germany, each measured from a server in that region.
Providers. Only services we call with our own paid API keys, under terms that allow publishing results. A provider whose terms require written consent before publishing is left out until we hold that consent.
Metrics. Time to first token and streaming throughput, with the number of requests, the run times and the raw per-request records published next to every summary figure. End-to-end timings are reported as measured. We do not subtract network handshake time to estimate a chip-only figure.
Cadence. Hourly runs at two prompt lengths. Failed requests are counted and shown, not dropped.
Write to [email protected]. For the pricing side, the Inference Calculator remains available.