Cartesia
Cartesia is a San Francisco company that sells speech models by API: text to speech (Sonic), speech to text (Ink) and a hosted platform for phone and web voice agents. On 4 Nov 2026 SuperDuperAi was accepted into its Startups Grant Program: a year of the Scale plan for free. This page covers what the company is, what the API does, how it compares, and where the grant goes in my products. Vendor numbers are marked as their claims.
Short version
- What we got. 12 months of Scale, 8M credits a month (the email says 88M characters a year, while the program page and 8M x 12 give 96M; I have not resolved the gap), plus $300 a month of agent credit. Scale concurrency, rollover up to 2x. The one condition: the Cartesia Startups logo on our website.
- What it is worth in speech. Their pricing page puts Scale at about 10,667 minutes of Sonic-3.6 audio a month, roughly 178 hours (vendor figure, ~750 credits per minute).
- The catch for us. Sonic speaks 44 languages including Russian and Turkish, but not Tatar or Bashkir. Ink-2, the transcriber, lists only English, French, Hindi, Japanese and Spanish. So Russian speech-to-text cannot go to Cartesia; speech out can.
- Where it goes (plan, nothing built yet). A voice for SuperVox, a voice line to my agents through the Meta glasses, voiceover for generated video, and client projects. Anything that hears a person stays on the device.
Who is behind it
Founded in 2023 by Karan Goel (CEO), Albert Gu (chief scientist), Arjun Desai and Brandon Yang, with Stanford professor Christopher Ré on the founding team (one source calls Ré only an advisor; TechCrunch lists him with the founders). Goel and Gu met as Stanford PhD students. Gu wrote the S4 paper (arXiv 2111.00396, with Goel and Ré) and, with Princeton’s Tri Dao, Mamba (arXiv 2312.00752). Cartesia’s pitch is that models of that family run faster and cheaper than transformers at inference.
Money raised, as reported: $22M seed led by Index Ventures in Dec 2024 (TechCrunch, which puts the total at $27M), a $64M Series A led by Kleiner Perkins in March 2025 (Fortune), and a $100M round with the Sonic-3 launch on 28 Oct 2025 (secondary sources only; I did not find Cartesia’s own announcement).
Why an SSM, in plain words
A transformer re-reads everything it has produced so far for each new token. A state space model (SSM) keeps one fixed-size summary and updates it as input arrives. For audio, which is a long stream where the next sample depends mostly on the recent past, a fixed cost per step is what makes “first sound in under 100 ms” plausible. TechCrunch’s phrase for the transformer side is rereading the whole book for every word.
graph LR
subgraph T[Transformer]
t1[new token] --> t2[attend over ALL previous tokens]
t2 --> t3[cost grows with length]
end
subgraph S[State space model]
s1[new token] --> s2[update one fixed-size state]
s2 --> s3[cost per step is constant]
end
This is the company’s thesis, not something I measured. Whether the 44-language Sonic is still a pure SSM is not stated in the docs I read.
Products as of 5 Nov 2026
| Product | What it does | Their claim or detail | Source |
|---|---|---|---|
| Sonic-3.6 (TTS) | Text to speech, 44 languages, emotion without SSML | ~90 ms to first audio byte; ids sonic-3.6, dated snapshot sonic-3.6-2026-08-27 |
docs |
| Sonic Turbo | Faster variant, smaller input per request | 40 ms model latency at Sonic 2 (Mar 2025); secondary source only | third-party summary |
| Ink-2 (STT) | Streaming transcription with built-in turn detection (turn.start, turn.eager_end, turn.end) |
en, fr, hi, ja, es; “lowest word error rate”, no numbers published | docs |
| Voice cloning | Instant clone from ~10 s of audio into 40+ languages; Pro clone on Startup and up | Pro clone training is free, speech from it costs the usual credits | docs |
| Managed Agents | Hosted voice agent: Ink + your LLM + Sonic, tools, telephony, versions | $0.03/min call time, $0.014/min telephony, LLM billed separately | docs, pricing |
| Line SDK | Python SDK for building voice agents in code (repo, 105 stars) | The docs index now has a “migrate from the Line SDK” page, so the managed route is the current one | docs |
| On-device | edge library, “on-device intelligence” (repo, 426 stars); a Sonic On-Device beta was announced in 2024 |
Docs index has no on-device page, so I treat it as not available to us | TechCrunch, GitHub |
| MCP server | text_to_speech, speech_to_text, voice clone/localize, pronunciation dictionaries, credit usage |
Hosted at mcp.cartesia.ai/mcp |
docs |
| SDKs | cartesia-python, cartesia-js, both Apache-2.0; cookbook and agent skills repos | GitHub |
For Claude Code the hosted MCP installs with one line:
claude mcp add --transport http --scope user cartesia https://mcp.cartesia.ai/mcp.
That route signs in through OAuth. The server also takes a plain API key as a Bearer token, which is what I use, because the browser sign-in fought with my automated Chrome. The key sits only in the macOS Keychain, and Claude Code asks a tiny script for the header on every connect (headersHelper), so the config file never holds it:
#!/bin/sh
# ~/.local/bin/cartesia-mcp-headers
k=$(security find-generic-password -s cartesia-api-key -w) || exit 1
printf '{"Authorization":"Bearer %s"}\n' "$k"
claude mcp add-json --scope user cartesia \
'{"type":"http","url":"https://mcp.cartesia.ai/mcp","headersHelper":"~/.local/bin/cartesia-mcp-headers"}'
Measured here on 5 Nov 2026: the account lists 9 Russian voices; a 5.8-second Russian sentence from sonic-3 over plain /tts/bytes came back with the first byte at 0.86 s and the whole file at 1.5 s. That is not the WebSocket latency they advertise, just what a one-shot HTTP call costs from a laptop.
Three ways to call TTS
graph TD
Q{Text arrives how?} -->|in pieces, from an LLM| W[WebSocket /tts/websocket]
Q -->|whole text, need timestamps over HTTP| E[SSE /tts/sse]
Q -->|whole text, audio only| B[Bytes /tts/bytes]
W --> W2[contexts and continuations, many utterances on one socket]
Per the endpoint comparison, every new HTTPS request pays TCP and TLS setup again, which the docs call often comparable to the audio’s own time to first byte. A WebSocket pays it once. Batch jobs (cached files, notifications) use Bytes.
Prices and competitors
Cartesia (pricing page, 5 Nov 2026): 1 credit per TTS character; STT 3 credits per second.
| Plan | $/month | Credits | TTS minutes (their est.) | Concurrent TTS | Overage per 1M |
|---|---|---|---|---|---|
| Free | 0 | 20K | ~27 | 2 | none |
| Pro | 5 | 100K | ~133 | 3 | $65 |
| Startup | 49 | 1.25M | ~1,667 | 5 | $45 |
| Scale | 299 | 8M | ~10,667 | 15 | $38 |
| Cartesia (Scale) | ElevenLabs | Deepgram | |
|---|---|---|---|
| TTS per 1K characters | ~$0.037 ($299 / 8M) | $0.04 Flash/Turbo (32 langs), $0.08 v3 (70+ langs) | $0.03 Aura-2 |
| STT per hour | $0.39 (Ink-2) | $0.22 Scribe v2, $0.39 Scribe v2 Realtime (90+ langs) | about $0.46 (Nova-3 multilingual streaming, $0.0092/min) |
| Agent per minute | $0.03 + LLM + telephony | not read | $0.050 to $0.163 all-in, depending on tier |
| Russian TTS | yes | yes | not verified |
Sources: Cartesia, ElevenLabs, Deepgram. I did not get OpenAI’s realtime prices (its pricing page did not render), so I do not compare it. Honest read: on list price the three are within a factor of two of each other for speech out. Cartesia’s case is latency and agent plumbing; ElevenLabs wins on language count and on STT price and coverage. For our grant the price is zero, so language coverage decides.
Glossary
- SSM (state space model). A sequence model that carries a fixed-size state instead of attending to all history. S4 and Mamba are the best-known ones.
- TTS / STT. Text to speech / speech to text.
- Time to first audio (TTFA, TTFB). Delay from the request until the first audio byte arrives. “Model latency” excludes the network; end-to-end includes it. Cartesia quotes 90 ms model and 190 ms end-to-end for Sonic-3 (secondary sources).
- WebSocket streaming. One open connection where text goes in and audio chunks come back as they are made, so speech starts before the sentence is finished.
- Continuation. Sending text to the same
context_idin pieces so intonation carries across chunks. - Turn detection. Deciding that a person has stopped talking, built into Ink-2 so no separate voice-activity detector is needed.
- Instant / Pro voice clone. A voice copied from seconds of audio / from a longer set, trained once and reused.
- Localization. Giving one voice an accent in another language (225 credits per accent).
Where the grant goes
Plans, nothing built yet. Four places:
| Where | What Cartesia does there |
|---|---|
| project-supervox | Voice for the call assistant: it speaks its hints and the agent’s answers instead of only showing them. Transcription stays as it is; Ink-2 has no Russian anyway. |
| The Meta glasses I bought | A voice line to my agents: I talk through the glasses, the agent answers in a Sonic voice. |
| Video generation | Voiceover for generated videos. |
| Client projects | Voice agents and voiceover for clients, inside the same 8M characters a month. |
The rule I keep from my privacy-first setup: people’s audio stays local, only text I have decided to speak goes to Cartesia. Cartesia says it may use customer data to improve its models unless you opt out (TechCrunch, 2024, not rechecked), so that setting gets checked before anything personal goes through.
What the dashboard shows after activation (5 Nov 2026): plan Scale, 8,019,636 model credits (the grant’s 8M plus what was left of the free tier), $300 for voice agents, renewal monthly. The meter: text-to-speech 1 credit per character, speech-to-text 3 credits per second, voice changer 15 credits per second; agent calls $0.03 a minute, telephony $0.014 a minute. Overages are off by default; if switched on they cost $38 per 1M credits.
What it sounds like
Four samples I generated on 5 Nov 2026 with sonic-3 over plain /tts/bytes, stock voices, no tuning. First byte came back in 0.46 to 0.81 s.
Russian, Alexei:
Russian, Tatiana:
English, Daniel:
Turkish, Leyla:
Terms and to-do
- The grant is credits on a Scale subscription activated through Stripe with a coupon (not recorded here). Cartesia may use our name and logo, and may ask for a social post; they offer a grant-recipient card with #BuildingWithCartesia.
- Done 5 Nov 2026: the Cartesia Startups logo is in the footer of superduperai.co, all six languages, under “Supported by”. That was the stated condition. Also set the overage switch off so nothing bills by accident.
- The fireside I attended (~28 Oct 2026, San Francisco) left no notes in the wiki beyond the date.
See also llm-providers-credits, tools-and-infrastructure, agent-landscape.