Cartesia Ships Sonic-3.6: A Streaming TTS Model That Now Leads Both Artificial Analysis Speech Arenas

Cartesia just put out Sonic-3.6, a new version of its real-time text-to-speech model, about three months after Sonic-3.5 shipped. The pitch this time is naturalness, and unlike most naturalness claims, this one is checkable against outside numbers. Sonic-3.6 now sits at #1 on both Artificial Analysis speech arenas, scoring 1,283 Elo on the Provider Voice board and 1,123 on the Controlled Voice board.

The second number is the one that actually tells you something. The Controlled board clones every competing model onto the same eight reference voices, which strips away any advantage from having a bigger or better voice catalog and leaves pure synthesis quality. Sonic-3.6 wins that comparison outright, with Sonic-3.5 landing second and ElevenLabs' Eleven v3 in third. So the improvement here is architectural, not cosmetic.

That architecture is unusual for the category: Cartesia builds Sonic on state space models instead of transformers, and it frames the standard speed-versus-naturalness tradeoff as something engineering choices can dissolve rather than something teams just have to accept. The practical payoff is latency — Cartesia states sub-90ms time-to-first-audio for Sonic, and 100ms transcript latency for its companion speech-to-text model, Ink-2. Both figures are vendor-stated model latency, not full round-trip numbers, so anyone building on this should benchmark their own pipeline rather than take the spec sheet at face value.

On the production side, Sonic includes features clearly built for agent transcripts rather than audiobook narration: inline tags for things like laughter, instant voice cloning from roughly 10 seconds of sample audio, custom pronunciation dictionaries with IPA overrides, and native handling of alphanumerics so order numbers and confirmation codes read correctly without preprocessing. Cartesia's demos also show Hinglish code-switching between Hindi and English, alongside more ordinary English speech with natural pauses and filler words.

Pricing tells its own story. Artificial Analysis puts Sonic-3.6 at $49 per million characters, half of ElevenLabs' Eleven v3 at $100, though still well above Speechify's Simba 3.2 at $10 for a 1,240 Elo score. Cartesia itself sells credits rather than raw characters — its Scale tier runs $299 a month for roughly 10,667 TTS minutes and 15 concurrent requests, with voice agent calls billed separately at 6 cents a minute. And it's worth flagging that this is a closed, hosted product: no open weights, no Hugging Face repo, and Sonic-3.6 is still in beta while Cartesia's own docs list 3.5 as the stable option.