Output Speed (tokens/s)

In plain words · How fast it writes — chunks of text produced per second

What does it measure?

How fast the AI writes its answer. For an answer of the same length, more tokens (roughly, chunks of characters) per second means less waiting.

Who evaluates it, and how?

Output speed measured per provider on real requests routed through OpenRouter, request-weighted within each day and reported as the median of the last 7 daily values (text-generation requests only). Our comparison tool also calls models through OpenRouter, so this is closest to the speed you feel on this site. Models with no new value for over 14 days, or not on OpenRouter, fall back to Artificial Analysis's 72-hour median under a standard workload (about 10k input tokens, at least 1,500 output tokens) — the two are measured under different conditions, so compare them with care. Reasoning-mode measurements are not used.

Reading the score

Unit is tokens/second, higher is faster. A model that thinks (reasons) at length can have a high number here yet start answering late, so read it together with 'time to first token'. The same model varies by provider and time of day.