Cartesia shipped Sonic-3.6, a streaming text-to-speech model the company says now ranks first on both of Artificial Analysis’s speech leaderboards, with under 90 milliseconds of time-to-first-audio.
The benchmarks
Sonic-3.6 scored 1,283 Elo on Artificial Analysis’s Provider Voice board and 1,123 on its Controlled Voice board, according to Cartesia’s own launch announcement. The Controlled board clones every competing model onto the same eight reference voices, isolating the synthesis engine itself from any advantage a larger voice catalog might provide, which makes it the more objective of the two rankings.
What’s under the hood
The model runs on state space models rather than the transformer architecture most competing TTS systems use, and generates expressive speech, including laughter, across more than 40 languages. It ships roughly three months after Cartesia’s previous release, Sonic-3.5, which remains what the company’s own documentation still lists as its stable model while Sonic-3.6 is in beta.
Why it matters
Sub-90ms latency is the detail that actually matters for real-time use, like a voice agent on a phone call, more than the naturalness scores that dominate most TTS marketing, since a technically excellent voice that lags is still unusable for a live conversation. Cartesia leading a benchmark it also helped define the terms of, via its own state-space-model approach, is worth a note of caution: Artificial Analysis is a third-party leaderboard, but a company’s own launch post is still the primary source citing these exact numbers here, not an independent re-test.
What to watch: independent latency and quality testing once Sonic-3.6 exits beta and becomes the documented stable model.
