Provider & model performance

Continuously sampled from TrustedRouter's monitor regions using a 7-day evidence sample, awaiting publication. Ranked by p50 time-to-first-token across all of a provider's models. Provider-attributed availability breaks latency ties. Different model mixes are not a controlled speed comparison.

Completion rate shows the end-user result. Provider availability excludes failures owned by TrustedRouter, customers, or route configuration, while still reporting those failures under their actual owner. Capacity acceptance measures requests not rejected for provider overload. Effective throughput is provider-reported output tokens divided by complete request time, so buffered delivery cannot inflate the result.

Models need 10 availability samples and providers need 30, each with at least 3 measured first tokens, to receive a rank. Thin availability and throughput measurements stay visible below the ranked set as warming. Unsupported route and probe-configuration rows are reported separately and do not count as provider downtime. No prompt or output content is ever stored.

This balanced evidence sample covers seven days. It is not today's availability or a complete traffic census.