Cartesia

Cartesia is a San Francisco-based voice AI company building real-time text-to-speech (and increasingly speech-to-text and full voice-agent) infrastructure around its Sonic model family. Its core technical differentiator is architecture: Sonic is built on state-space models (SSMs) rather than the transformer/diffusion stacks most competitors use, which is what lets Cartesia advertise time-to-first-audio in the 40-90ms range. The company is squarely a developer-API play, not a no-code product. You get a low-latency TTS endpoint with WebSocket streaming, instant and professional voice cloning, 40+ languages, custom pronunciation dictionaries, and emotion/laughter controls. You do not get a packaged business tool that wires itself into your helpdesk. As one review put it, Cartesia is "an excellent engine, but you still need to build the car."

Where Cartesia wins is raw speed and the real-time conversational use case: voice agents, IVR/telephony, and live dubbing where every 100ms of latency degrades the experience. Independent third-party numbers back the speed claim directionally, the Coval benchmark (May 2026) clocked Sonic-3 at 188ms P50 TTFA, faster than ElevenLabs Turbo/Flash v2.5 and Deepgram Aura-2, though it's worth noting that real measured latency is meaningfully higher than the ~40ms marketing figure (which is time-to-first-byte under ideal conditions), and Cartesia's latency variance (a 100ms inter-quartile range in that test) is wider than some rivals, which matters for production consistency. On pure quality, Cartesia is competitive but no longer dominant: on Artificial Analysis's Elo leaderboard Sonic 3.5 briefly held #1 before being overtaken, and it now trails newer entrants like Inworld on naturalness Elo.

Commercially, Cartesia is priced aggressively for developers with a credit model (1 credit/character for TTS) and a tiered ladder from a free 20k-credit plan up through Pro ($5), Startup ($49), Scale ($299) and custom Enterprise, plus per-minute voice-agent and telephony pricing. It carries enterprise compliance (SOC 2, HIPAA, GDPR, PCI) and supports on-prem/on-device deployment, which broadens its appeal beyond startups. The main caveats: there's little independent review-site presence (no real G2/Trustpilot/Capterra footprint), self-reported logging/observability is a known weak spot, and because it's an infrastructure primitive rather than an agent platform, teams must build and maintain the surrounding orchestration themselves.