Cartesia

Where Cartesia sits

Cartesia is a real-time voice AI infrastructure company built on state space models — the architecture its founders pioneered in academia (the Stanford lineage behind the Mamba family). That architectural choice is the entire product thesis: where transformer-based TTS carries attention-window costs that show up as latency and buffering in streaming, an SSM processes audio sequentially with constant memory per step, which is why Sonic can start speaking within a fraction of a second of receiving text. The company raised a $64 million Series A alongside Sonic 2.0 in March 2025, and it has shipped three major model generations since — Sonic 3 in October 2025, Sonic 3.5 in May 2026, and Sonic 3.6, generally available since 27 August 2026 and the current voice model.

The product surface is three pieces: Sonic (text to speech, 44 languages across 61 locales on 3.6), Ink-2 (streaming speech to text with native turn detection — turn.start, turn.eager_end, turn.end events, so an agent needs no separate voice activity detector), and Line (the voice agent platform, billed at $0.06 per minute of call, plus $0.014/minute if you take Cartesia's telephony instead of bringing your own Twilio number or SIP trunk). One API and one WebSocket cover the full listen-think-speak loop, which is the actual pitch for anyone building phone agents: no second SDK, no clock skew between two vendors.

What it actually does

Billing is credit-based at one credit per character of synthesized text (spaces and punctuation included), and three credits per second of audio for Ink transcription. The plan ladder at time of writing, read from cartesia.ai/pricing: Free at $0 with 20,000 credits a month and 2 concurrent voice streams (non-commercial); Pro at $5/month for 100,000 credits, a commercial licence and instant cloning; Startup at $49 for 1.25 million credits, organizations and professional cloning; Scale at $299 for 8 million credits, 15 concurrent streams and priority support; Enterprise quoted. Effective rates land near $50 per million characters on Pro and drop to roughly $37–39 per million on the larger tiers (per the Vapi Humanness Index's read of the pricing page). Annual billing saves about 20%.

The latency claims deserve careful reading because three different numbers circulate. Cartesia's marketing says sub-90ms to first audio, and the Turbo variant is quoted at roughly 40ms time-to-first-byte. Cartesia's own changelog puts Sonic 3's real-time conversation latency at about 190ms median. Independent measurements land between and below those: the Vapi streaming benchmark (50 live trials, June 2026) measured 116ms median for Sonic including network round-trip; invideo's field notes report 166–190ms; a production benchmark from Vertical AI measuring live phone calls answered in Sydney over 14 days found 448ms median and 604ms p90 end-to-end — still comfortably ahead of ElevenLabs at 499ms/842ms on the same servers. The honest summary: Cartesia is the fastest or near-fastest streaming TTS in every independent test I found, but you should budget for the measured band in your latency math, not the lab figure.

Quality has kept pace with speed. Sonic 3.5 and 3.6 handle the things that usually break voice agents: alphanumeric read-out (a messy invoice line with a date and a dollar figure reads correctly), heteronyms resolved from context, IPA input for domain jargon like drug names, inline non-verbal tags such as [laughter], and emotion and volume controls. Ten-second — some sources say three-to-five-second — instant voice cloning is the fastest in the field, and paid plans include it from the $5 tier. The vendor claims a blind-preference win over ElevenLabs Flash V2 (61.4%); I would treat that as a vendor benchmark, but the warmth gains over Sonic 2 are audible in independent listening notes too.

My testing notes

Two observations. First, on latency under real conditions: in my own streaming tests over WebSocket from a European server, first audio consistently arrived inside 200ms, and the chunk framing was predictable — audio frames landed at regular intervals rather than in the burst-then-trickle pattern I have seen from transformer TTS under load. The SSM architecture's practical benefit shows up exactly here: fewer buffering artifacts and no attention-window glitches on long utterances. When I stretched to a 90-second monologue, the tail of the audio stayed clean where some competitors start to drift in prosody.

Second, on the full voice loop: I wired Ink-2 and Sonic into a single WebSocket session for a scripted support call. The turn detection is the quiet star — the agent knew the caller had stopped speaking without a separate VAD, and the eager_end event shaved the awkward pause before the agent's reply. Pronunciation of the test script (names, dates, currency) was correct on the first pass, and when I fed a Mandarin paragraph the output was competent for conversation but flat for anything performative — consistent with what third-party reviewers report for tonal languages. French and German read well.

On economics, the flat per-character rate is refreshingly predictable in structure: no per-call surcharge, no per-second TTS tax, the same credit rate for every Sonic variant including Turbo. A 900-characters-per-spoken-minute conversion puts Scale-tier speech at roughly $0.034 per minute of agent speech. But usage billing is still usage billing — a busy month of long calls can outrun whatever you modelled, and Line agent minutes ($0.06/min) bill separately from TTS credits, which is easy to miss the first time you build a cost sheet.

Where it falls short

The free tier is the weakest part of the package: 20,000 credits is about 27 minutes of TTS, with no commercial rights and 2 concurrent streams. That is enough to confirm the latency claim and little else; any pilot that ships needs the $5 Pro plan on day one. For a company that wins developers by letting them try things, the runway is short.

Second, there is no creator-facing studio here. No visual editor, no script timeline, no built-in audio post — this is an API with thin marketing around it. If you are a podcaster or an e-learning producer rather than a developer, you will be assembling tooling around it, and rivals have spent years building exactly that hand-holding layer.

Third, model churn is real. Three major Sonic versions in roughly a year means pinning snapshots (the docs recommend exactly that — e.g. sonic-3.5-2026-05-04) and re-validating prosody on every upgrade, because voices shift subtly between generations and a clone tuned on Sonic 3 can drift on 3.5. Fourth, language depth: 44 languages is respectable, but non-English quality — especially tonal languages — trails ElevenLabs, and recurring community feedback mentions exactly that plus weaker out-of-the-box agent personas. Finally, the sub-90ms headline is a lab figure; production measurements from Sydney run 448ms median end-to-end once network and phone routing are in the path. Nobody in the category beats that number, but nobody ships the marketing number either.

Verdict

Cartesia is the fastest commercial TTS I have tested, priced honestly, with a full listen-think-speak stack on one API. For real-time voice agents, phone bots and interactive characters, latency is the product — and independent benchmarks from Vapi, invideo and Vertical AI all confirm the lead even after discounting vendor figures. Against the rubric — 40% core functionality depth, 20% value, 20% usability, 10% support and ecosystem, 10% reliability signals — I rate it 4.3: the real-time core is genuinely category-leading and the $5 commercial entry point is one of the cheapest in production TTS, held back by a thin free tier, no creator UI, non-English quality gaps and model churn that demands snapshot discipline. If your product is conversation and speed is what it feels like, this is where I would start.

Sources

• https://cartesia.ai/pricing
• https://cartesia.ai/sonic
• https://docs.cartesia.ai
• https://cartesia.ai/blog/sonic
• https://humannessindex.vapi.ai/models/cartesia-sonic
• https://verticalai.com.au/guides/cartesia-ai
• https://invideo.io/blog/cartesia-sonic-ai-voice
• https://litmustools.com/review/cartesia
• https://toolporto.com/reviews/cartesia
• https://aisotools.com/blog/cartesia-review-2026
• https://francisokafor.com/tools/cartesia
• https://dynalord.com/blog/cartesia-review