Cartesia review
The fastest voice API for real-time agents
A genuinely strong, actively developed TTS/voice API that wins on latency and price for real-time voice agents. If speed on live calls is your bottleneck, it is the front-runner; for general content narration the field is more crowded.
Best for: Developers building latency-critical, real-time voice agents — phone bots, support lines — who want TTS and STT from a single vendor. · From: Free tier $0 (~27 TTS min/mo + ~1h51m STT, 1 agent slot, personal/non-commercial). Paid from $5/mo Pro ($4/mo annual — adds commercial license + instant voice cloning); Startup $49/mo (professional voice cloning, organizations); Scale $299/mo (priority support, high concurrency); Enterprise custom. Live voice-agent calls $0.06/min across all tiers; telephony via a Cartesia-provided number $0.014/min. (Eligible startups: 12 months of Scale free via the AWS offer.)
Pros
- + Class-leading latency (~40ms time-to-first-byte on Sonic-3.5 Turbo, sub-90ms on the standard model) — the whole point for live phone and voice agents
- + Instant voice cloning from ~10 seconds of audio (paid tiers), plus professional cloning, across 40+ languages
- + Ships both the TTS model (Sonic-3.5) and an STT model (Ink-2, with built-in turn detection), so the full voice pipeline runs from one vendor
- + Aggressive entry pricing: Pro is $5/mo and live voice-agent calls run $0.06/min
- + Inline tags for laughter and non-verbal sounds make output genuinely lifelike, not just clear
Cons
- − Free tier is thin and non-commercial (~27 TTS minutes/month, 1 agent slot) — enough to prototype, but a commercial license and voice cloning both require the $5 Pro plan
- − Human/priority support is gated behind the $299/mo Scale plan; lower tiers are self-serve
- − Narrower toolset than ElevenLabs for content work (no deep dubbing or long-form studio)
- − Credit, minute, and per-call/telephony accounting takes effort to forecast at scale
Setup
Cartesia is API-first. You grab a key from the dashboard and are calling the REST or WebSocket endpoint within minutes; there are official SDKs and first-party integrations with the common voice-agent frameworks (LiveKit, Vapi, Pipecat). The streaming WebSocket is the path you actually want for live calls, and it is well-documented. No credit card is needed to start.
The AI / voice quality
The flagship TTS model is Sonic-3.5, and the pitch is speed: roughly 40ms time-to-first-byte on the Turbo variant and sub-90ms on the standard model — genuinely class-leading for real-time synthesis. Voices are natural and expressive, and Sonic-3.5 adds inline control tags for laughter and non-verbal sounds that make output sound human rather than merely clear. It covers 40+ languages (42) at native quality, with instant voice cloning from about ten seconds of audio on paid tiers and professional cloning from the Startup plan up. Cartesia also ships its own STT model, Ink-2, with built-in turn detection — so you can run the whole listen-think-speak loop from one vendor instead of bolting Deepgram or Whisper onto a separate TTS.
Pricing
Free is $0 (20K credits, ~27 TTS minutes plus ~1h51m STT, one agent slot), but it is personal/non-commercial — you need Pro for a commercial license. Pro is $5/mo ($4 annual) with instant cloning and three slots; Startup is $49/mo (professional cloning, organizations, five slots); Scale is $299/mo (priority support, higher concurrency); Enterprise is custom. Live voice-agent calls are billed $0.06/min across every tier, and telephony through a Cartesia-provided number adds $0.014/min. Eligible startups can get 12 months of the Scale tier free through the AWS offer.
Free / cheaper alternatives
ElevenLabs has richer content tooling (dubbing, long-form studio, a bigger voice library) if narration — not latency — is your priority. For zero-cost or self-hosted, Piper and Kokoro run locally with no per-minute fee, trading some naturalness and easy cloning for control and privacy. On STT specifically, Deepgram is the incumbent worth benchmarking Ink-2 against.
Verdict
If your bottleneck is speed on live phone or voice-agent calls, Cartesia is the front-runner: lowest-in-class latency, a single-vendor TTS+STT stack, and aggressive entry pricing at $5/mo. For general content narration the field is more crowded, and ElevenLabs is often the safer pick.