Cartesia Review (2026): Pricing, Features & Honest Verdict

TLDR

Cartesia builds Sonic, the lowest-latency realtime text-to-speech model on the market, plus Ink transcription and Line voice agents, all on state-space architecture. If you are building voice agents where every millisecond counts, this is the leader. Best for: engineering teams shipping realtime voice. Price: Free plan: yes, paid from $4/mo. Rating: 8.3/10.

What is Cartesia?

Cartesia is a realtime voice AI company. Its flagship product, Sonic, is a text-to-speech model that generates natural-sounding speech with a time-to-first-audio of roughly 40 milliseconds, which is the kind of latency you need for a voice agent that talks back without an awkward pause. Around Sonic sit two more products: Ink, a streaming speech-to-text model for fast transcription, and Line, a platform for building and shipping voice agents on top of both. Together they cover the full voice loop: listen, think, speak.

The company was founded by Karan Goel, Albert Gu, Arjun Desai, Brandon Yang, and Chris Re, a group of PhDs who met at the Stanford AI Lab and invented state-space models, the architecture behind Mamba. That pedigree matters here. State-space models are efficient at long-context, streaming workloads in a way transformer-based speech models are not, which is exactly why Cartesia can hit the latency numbers it does. Cartesia raised around $191M total, including a $100M round led by Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA, and it shipped the Sonic-3 generation of models off the back of it. We dug into Cartesia because voice agents are one of the few AI categories where speed is a hard product requirement, and Cartesia is the team that built the speed advantage into the model itself rather than bolting it on after the fact.

What Are Cartesia’s Key Features?

Sonic: lowest-latency text-to-speech

Sonic is the reason most teams show up. It posts a time-to-first-audio around 40 milliseconds and sub-90ms latency overall, which is fast enough that a voice agent feels like it is responding, not buffering. Sonic-3.5 supports 42 languages including English, Spanish, French, German, Hindi, and Japanese, and it ranks at or near the top of independent naturalness rankings. For realtime use, that combination of speed and quality is the whole game. Most TTS models force a choice between fast-but-robotic and natural-but-slow; Sonic is the rare one that does not make you pick. The Sonic-3 generation also added more expressive controls, so an agent can carry emotion and emphasis instead of reading in a flat monotone.

Instant and professional voice cloning

Sonic can clone a voice instantly from about 10 seconds of audio, available from the $4 Pro plan up. If you need a higher-fidelity custom voice, professional voice cloning unlocks on the $39 Startup plan after a one-time training step. This is how you give an agent a consistent brand voice instead of a generic stock read.

Ink: streaming speech-to-text

Ink is Cartesia’s transcription model, built for the same realtime constraint as Sonic. Ink-2 handles streaming transcription so a voice agent can understand a caller as they speak rather than waiting for them to finish. Pairing Ink with Sonic in one stack means you are not stitching together two vendors with mismatched latency budgets, which is a common source of lag when teams glue a third-party transcriber onto a separate TTS provider. One vendor, one latency profile, one bill.

Line voice agents and flexible deployment

Line is the agent layer: it lets you build phone and voice agents on top of Sonic and Ink, with calls billed at $0.06 per minute and Cartesia telephony at $0.014 per minute. Underneath all of it, the state-space architecture supports cloud, on-premise, and on-device deployment with in-region processing, so teams with data-residency or compliance rules are not forced into a single cloud.

How Much Does Cartesia Cost?

Cartesia starts free, with roughly 27 minutes of Sonic TTS and about 1 hour 51 minutes of Ink transcription per month, capped at 2 concurrent TTS requests and no commercial-use license. Paid tiers climb from there: Pro at $4/mo adds commercial use, instant voice cloning, and 3 concurrent requests; Startup at $39/mo jumps to roughly 1,667 minutes of TTS, professional voice cloning, and organizations; Scale at $239/mo gives you about 10,667 minutes, priority support, and high concurrency. Enterprise is custom and adds DPAs, BAAs, SSO, and a shared Slack channel.

The pricing model is consumption-based: Sonic bills at roughly 1 credit per character. That is cheap at moderate volume and gets pricey at high volume, which is the honest trade-off. By comparison, ElevenLabs uses a subscription-plus-credits model where Creator runs $22/mo and Pro runs $99/mo, and Azure AI Speech bills neural voices at about $15 per million characters with cheaper committed tiers. Cartesia tends to win on latency and per-minute call economics; the bulk character providers can win on raw price if you are generating huge volumes of non-realtime audio.

Model your numbers before you scale. For a voice agent doing thousands of minutes of live calls a month, Cartesia’s $0.06 per minute call pricing and low latency are the right fit. For batch-generating millions of characters of narration with no latency requirement, a flat-rate character provider may come out cheaper. Match the pricing model to your workload. The jump from Startup at $39/mo to Scale at $239/mo is also worth planning for: it is a roughly 6x price step that buys roughly 6x the minutes plus high concurrency, so the per-minute rate stays consistent, but you want to know which tier your call volume lands in before launch rather than after a surprise overage.

Who is Cartesia Best For?

Use Cartesia if you are an engineering team building realtime voice: phone agents, interactive assistants, in-app voice features, or anything where a half-second delay kills the experience. The 40ms time-to-first-audio is a genuine technical edge, and the founding team built it into the architecture rather than optimizing around a slow model. The free tier and $4 Pro plan make it easy to prototype before you commit.

Skip Cartesia if you mostly need batch voiceover or narration with no latency requirement, where cheaper bulk character pricing wins, or if you want a polished no-code studio for non-technical users. Cartesia is a developer-first API. If you need a deep catalog of pre-made character voices and languages out of the box, ElevenLabs still has the larger library.

Best Cartesia Alternatives

ElevenLabs

ElevenLabs is the quality and voice-library leader, with the deepest catalog of voices and languages and strong professional cloning. Its Creator plan is $22/mo and Pro is $99/mo, using a subscription-plus-credits model. ElevenLabs wins on sheer voice variety and a friendlier studio; Cartesia wins on realtime latency and per-minute call economics. If you are building live voice agents, Cartesia is faster. If you want the widest range of ready-made voices, ElevenLabs leads.

Azure AI Speech

Azure AI Speech is the enterprise workhorse, billing neural voices at about $15 per million characters with committed tiers that drop toward $7.50 per million at high volume. It is the cheap, dependable option for organizations already on Azure that need huge batch volumes. It does not match Cartesia on realtime latency or voice naturalness, but for non-realtime bulk synthesis on a budget, it is hard to beat on price.

KugelAudio

KugelAudio is a newer realtime TTS API built around European data sovereignty, with pay-per-use rates from about EUR 0.043 per minute and optional on-premise hosting. It is the option to look at if your hard requirement is that voice data never leaves the EU. It has a thinner track record than Cartesia and no free tier, so it is a fit for GDPR-driven teams specifically rather than a general-purpose replacement.

Final Verdict: Is Cartesia Worth It?

Cartesia earns a high rating because it leads on the one metric realtime voice cannot fake: latency. A 40ms time-to-first-audio combined with top-tier naturalness, built by the team that invented state-space models, is a real and defensible advantage. The free tier and $4 Pro plan make it cheap to try, and flexible deployment covers compliance-heavy teams. For voice agents, phone systems, and interactive apps, this is the model we would reach for first.

It is not perfect. The per-character credit pricing scales up faster than a flat enterprise contract at very high volume, and the voice and language library is narrower than ElevenLabs. If your workload is batch narration with no latency requirement, a bulk character provider may be cheaper. But if you are building anything where the AI has to talk back in real time, Cartesia is the leader, and it is worth building on.