Artificial Analysis Output Speed

Artificial Analysis's live-measured output speed for a model's API: tokens generated per second, a performance measure, not a correctness or quality score.

Also known as: Artificial Analysis Speed Index, AA Output Speed, Output Tokens per Second

unassessed

What it measures

This page documents Artificial Analysis's Output Speed metric — the throughput figure the publisher headlines under "Speed" on its own site — as this repository's artificial_analysis_speed_index. It is a pure performance measurement, not a capability or correctness score: Artificial Analysis sends a live prompt to a model's public API and times how many tokens per second stream back after the first token arrives. Unlike the Intelligence Index, Artificial Analysis does not publish a single blended "Speed Index" combining multiple timing metrics into one number; Output Speed is reported alongside, not merged with, separate metrics for time to first token and end-to-end response time. Workloads vary by input length (about 1,000, 10,000 or 100,000 input tokens, plus a vision workload of one megapixel image and roughly 1,000 text tokens) because both time-to-first-token and output speed itself shift with prompt length and technique such as speculative decoding.

Task format

Live API calls: Artificial Analysis sends standardized, freshly generated prompts of a fixed input-token length to a model's public endpoint and streams the response, timing token arrivals to derive tokens-per-second and latency figures.

Models reporting this benchmark

Data

This page as JSON · Edit on GitHub