GradiumBENCHMARKS

Data read on November 19, 2026

How these numbers are read

What is measured

Coval sends the same text to every model's streaming API from US East and records when the first audio byte arrives (time to first audio), how much silence precedes speech, and transcribes the audio to compute a word error rate.

Median and P95

The median is the typical turn. The P95 is the slow turn: 5% of requests are slower. Voice agents feel the P95, because one slow reply in twenty is enough to break a conversation.

Leading silence

Some models return audio quickly but start with silence. Coval measures that silence separately, so a fast first byte that starts with silence does not count as fast speech.

Refresh

This site reads Coval's public leaderboard again every night and shows the date of the read on every page. Coval's figures move from day to day; the date tells you which read you are looking at.

What we never do

No figure on this site is estimated, rounded up or combined into a score. When a source does not publish a measure, the page says so.

Sources