Cartesia Review

90ms at P50 is the number that matters. Voice agents break when TTS lags, and Cartesia's state-space architecture — not transformer-based — is specifically why they're faster than ElevenLabs at this percentile. That's a real technical moat, not a marketing claim.

Two things give me pause. One: Enterprise is the only tier without a number, so the compliance package we'd actually need starts with a sales call. Two: credit-based billing tracks traffic, not headcount, so year-two spend is a forecast rather than a line item. The free tier gets you started, but I'd want a signed SLA before betting a production voice agent on them.

The Instant Voice Cloning from a 10-second sample, 40+ language coverage, and the browser playground make this fast to evaluate. Pilot it in a contained use case — healthcare triage bot, customer support IVR — before you standardize.

Emotion via markup tags — laughter, excitement, sadness — plus context-aware acronym handling shows someone thought hard about expressive fidelity, not just phoneme accuracy. That's the difference between a voice render engine and a voice design system. The instant clone from a 10-second sample, plus a Pro Voice Clone tier for enterprise fine-tuning, gives brand teams two distinct levers depending on budget and polish requirements.

The tradeoff worth naming: Cartesia is an API-first developer tool. There's no visual voice design workflow, no waveform editor, no asset library with version history. A Creative Director needs engineering mediation for every iteration. ElevenLabs has a more self-serviceable interface for non-technical brand stakeholders.

For regulated industries — healthcare, finance — SOC 2 Type II plus HIPAA plus PCI Level 1 in one stack is genuinely rare. If the product roadmap includes voice agents in those verticals, this compliance posture eliminates a year of procurement friction.

Sonic-3's latency story is specific: under 100ms model latency, one customer citing 90ms, and a claim of 4x faster than the next competitor. That's a number you can put in a build-vs-buy memo. SOC 2 Type II, HIPAA, and PCI Level 1 certification makes regulated-industry deployment viable without a compliance detour.

The rate card is public. Five tiers: $0, $5, $49, $299, Enterprise custom. Startup at $49 buys roughly 1,667 TTS minutes — about $0.029 a minute. Scale at $299 buys 10,667 minutes, near the same rate. ElevenLabs publishes $0.10 per 1,000 characters. Different denominator, so compare on minutes.

Instant voice cloning starts at $5, not on the free plan. Professional clones start at $49 — two of them, four at $299. Voice agent calls run $0.06 a minute, telephony $0.014. I couldn't find an overage rate for credits past plan. Contract terms aren't posted. Procurement will still make one call.

90ms P50 latency is the number that matters here. ElevenLabs sounds warmer in isolation, but when you're building a voice agent that needs to respond inside a natural conversational pause, Cartesia's 4x speed advantage over its next competitor isn't a spec sheet brag — it changes what's architecturally possible. Streaming output means your pipeline doesn't wait for full render. That's a different class of tool.

The emotion markup tags and laughter generation via Sonic-3 are genuinely useful for anything beyond flat IVR delivery — companion apps, gaming NPCs, healthcare intake flows. Instant voice cloning from a 10-second sample handles quick prototyping. Pro Voice Clone for production. That two-tier structure makes sense. The browser playground lets you hear configurations before touching the API. Good sequencing.

The gap: billing runs on credits, not minutes. The $49 Startup plan lists about 1,667 TTS minutes, so you're converting to per-minute yourself every time you size a workload. That's a small tax on every budget conversation. SOC 2 Type II and HIPAA compliance are table stakes for healthcare — they have them, which matters — but the enterprise plan is fully custom. Day-3 reality: prototype on the $5 tier, model the credit burn before you commit a production workload.

The number that matters here is 100ms. That's the latency ceiling Sonic-3 is designed to stay under, and one customer reportedly hit 90ms. For a live voice agent — the kind where a human expects an answer, not a pause — that gap between Cartesia and its next best alternative feels like the difference between a conversation and a conference call. The playground-in-browser is a smart call too. Test your script, tweak the emotion tags, hear laughter or excitement rendered in 40-plus languages, all before touching a line of code.

The tradeoff is that the plans are counted in credits, not minutes. There's a free tier and $5 gets you voice cloning, but you're doing arithmetic to work out what an hour of audio actually costs. Fine for solo devs experimenting, less fine if you're running a procurement process.

This is an API product dressed up in a clean web interface — mobile parity isn't really the point. If your team needs instant voice cloning from a 10-second sample and the enterprise compliance to deploy it in healthcare or finance, Cartesia earns the look. Otherwise it's a developer tool, full stop.

Three things I check first. One: the latency claim is P50-to-P99 across global regions, not just a cherry-picked datacenter result. Two: SOC 2 Type II, HIPAA, and PCI Level 1 in the same package is unusual for a startup — that's real compliance work. Three: '40+ languages covering 95% of the world' is the kind of superlative that ages poorly, but 9 named Indian languages is specific enough to be checkable.

The instant voice clone from a 10-second sample is a crowded feature — ElevenLabs does this too. The differentiation is latency and the state-space architecture behind it. One customer citing 90ms specifically is a real number, not a range. Could be cherry-picked. Watch whether that holds at scale.

Pricing is on the page: $0, $5, $49, $299, Enterprise custom. What I couldn't find is the overage rate once a plan's credits run out. Exit portability is decent — REST API, multi-language SDKs — but proprietary emotion markup tags create some lock-in friction.