Cartesia Sonic 3.6

Cartesia Sonic 3.6 is a text-to-speech model using a State Space Model architecture that scales linearly with sequence length, unlike transformer-based models whose inference grows quadratically. This architecture enables sub-90ms time-to-first-audio at production load, with P90 first-audio latency measured at 351ms in the full Cartesia voice agent pipeline by independent benchmarks. The model covers 44 languages across 61 locales and was released August 27, 2026. Combined with Ink-2, Cartesia's streaming speech recognition model, Sonic 3.6 forms a complete voice stack for real-time agent applications. Managed agent calls cost $0.03 per minute, with an optional Cartesia-provided phone number adding $0.014 per minute. Developers supply their own LLM while Cartesia handles speech recognition and synthesis. Voice cloning is available from approximately 10 seconds of reference audio, enabling rapid deployment of custom or branded voices. The free tier provides 20,000 TTS credits per month. Cartesia Sonic 3.6 is optimized for high-volume deployments where the speech synthesis layer is a primary cost driver. At $0.03 per minute, operators running hundreds of thousands of simultaneous minutes benefit from the linear scaling of the SSM architecture and consistent sub-90ms TTS latency. Python and TypeScript SDKs plus a streaming WebSocket API support direct integration into existing voice infrastructure.