Sonic-3.6 Hits 1,283 Elo and Leads Both Artificial Analysis Speech Arenas
Cartesia released Sonic-3.6 on August 18, 2026, roughly three months after Sonic-3.5, and it immediately claimed the top position on both Artificial Analysis speech leaderboards — 1,283 Elo on the Provider Voice board and 1,123 Elo on the Controlled Voice board. The Controlled Voice result is the more meaningful number: that board clones every competing model onto the same eight reference voices, isolating synthesis quality from catalog breadth. Sonic-3.6 leads it, Sonic-3.5 places second, and ElevenLabs Eleven v3 sits third.
Architecture and Latency
Sonic runs on state-space models rather than transformers — an architectural choice Cartesia frames as structural rather than a tuning tradeoff. SSMs process a transcript as a continuous stream rather than buffering a full chunk before synthesis begins, which is why Cartesia states sub-90ms time-to-first-audio. The companion Ink-2 speech-to-text model posts 100ms transcript latency with native turn detection. Both figures are vendor-stated model latency, not end-to-end round-trip measurements over a production network; developers building voice interfaces — the kind increasingly discussed as agentic-era data pipelines evolve — should benchmark their own paths before committing to an SLA.
The SSM design surfaces in real latency budgets in a way transformer-based chunked models do not. A chunked model must complete a meaningful audio segment before emitting anything; an SSM can push the first audio chunk out before the input sentence is complete. That gap matters acutely in inbound support agents and IVR-replacement deployments where perceived responsiveness determines whether a caller stays on the line.
Production Feature Set
Sonic-3.6 exposes controls designed for agent transcripts rather than narration. Inline expression tags such as [laughter] go directly in the transcript text. Pronunciation dictionaries accept IPA overrides — the source gives <<s|ə|ˈ|p|i|n|ə>> for "subpoena" as a concrete example. Speed and volume are API parameters; through the LiveKit Agents plugin, speed is valid from 0.6 to 1.5 and volume from 0.5 to 2.0 on Sonic-3 models. Order numbers, phone numbers, and confirmation codes are read natively without preprocessing. Voice cloning requires approximately 10 seconds of reference audio. Launch demos include Hinglish code-switching between Hindi and English, relevant for contact-center deployments targeting South Asian markets.
Pricing
| Model | Controlled Voice Elo | Price per 1M Characters |
|---|---|---|
| Cartesia Sonic-3.6 | 1,123 | $49.00 |
| ElevenLabs Eleven v3 | 3rd place | $100.00 |
| Speechify Simba 3.2 | 1,240 | $10.00 |
Artificial Analysis normalizes Sonic-3.6 at $49.00 per 1M characters — half of ElevenLabs Eleven v3 at $100.00 and well above Speechify Simba 3.2 at $10.00, which posts 1,240 Elo. Cartesia sells capacity in credits rather than raw character counts. The Scale plan at $299/month includes approximately 10,667 TTS minutes and 15 concurrent requests. Line voice agents bill separately at $0.06 per minute, with Cartesia-provided telephony adding $0.014 per minute on top. The Free tier carries no commercial license; commercial use begins at the Pro tier at $5/month. Concurrency caps — 2, 3, 5, and 15 simultaneous TTS requests on Free, Pro, Startup, and Scale respectively — can become the binding constraint during call surges well before minute quotas are exhausted.
Availability
Sonic-3.6 is in beta on Cartesia's hosted API. There are no open weights and no Hugging Face repository; the model is closed and commercial. Cartesia's own documentation still lists Sonic-3.5 as stable, and distribution partners carry 3.5. Developers integrating today should pin explicitly to 3.6 and track the stable-channel promotion separately.
The Controlled Voice leaderboard result is the figure that will move procurement conversations. It demonstrates that Cartesia improved the synthesis engine itself — not merely expanded a premium voice catalog — which is exactly the kind of independently verifiable claim that distinguishes a substantive model release from a marketing refresh. Combined with the SSM latency profile and a per-character price below the incumbent leader, Sonic-3.6 sets a new cost-quality reference point for production voice AI.