Latency and speed from client traffic, uptime from a probe every minute.

Aggregates only, computed from live production metrics: no request contents, no client identifiers — we run zero data retention, so there is nothing to expose. TTFT, generation speed and the success rate come from client traffic; uptime and probe latency come from a health probe we send every minute, and the hourly strip combines that probe with errors on client traffic. TTFT is measured in our proxy on the GPU node, from the moment the request is read to the first streamed byte, prefill included. Latency instrumentation live since Aug 23, 2026; serving production traffic since Aug 22, 2026. Page data refreshes every 5 minutes.