What it is
Sonic 3.6 became generally available on 27 August 2026 (snapshot sonic-3.6-2026-08-27). Cartesia says it is fully backwards compatible with Sonic 3.5, keeps professional voice clones made on older models, and adapts pacing, pauses and intonation to the transcript without SSML tags.
It speaks 44 languages, adding Odia and Urdu, with better Hinglish and Indian names and places. Cartesia offers dated snapshots that never change, a rolling sonic-3.6 alias that tracks the latest stable snapshot, and a sonic-preview channel that can change without notice.
Our review
Sonic 3.6 is fourth in Artificial Analysis's speech arena at 1289 Elo, behind Eleven v4 Turbo (1335), Eleven v4 (1329) and Alibaba's Qwen-Audio-3.1-TTS-Plus (1301). That is a 98-point jump from Sonic 3.5 at the same $49 per million characters, and it puts Cartesia 11 points ahead of Gemini 3.8 Flash TTS.
On price it sits in an awkward middle. Eleven v4 Turbo is ranked higher and costs $40 per million characters on Artificial Analysis's comparison, while Gemini 3.8 Flash TTS ($16.5) and Qwen-Audio-3.1 ($19.3) cost well under half as much for scores within about 12 points. Cartesia's case rests on its agent plumbing instead: a WebSocket API with context IDs for streaming partial transcripts, word and phoneme timestamps, telephony encodings such as mu-law at 8 kHz, and a hosted agent platform at $0.06 a minute.
The versioning is the strongest reason to choose it for production. Dated snapshots never change after release, so a contact-centre voice can be frozen while the team evaluates the next one, and Cartesia publishes no retirement dates for older snapshots. Latency is another matter: the docs call Sonic the fastest model but give no time-to-first-audio figure, so latency has to be measured on your own traffic.
The data terms need attention before a regulated deployment. Cartesia's privacy policy, last revised in June 2024, says it may use uploaded text, voice recordings and generated audio to train its models, with an opt-out form that applies only from the date it is filed. DPAs, BAAs, SSO and on-premise or VPC deployment are Enterprise-only, while SOC 2 Type II covers the whole company.
Where it earns its place
A voice-AI startup running an outbound appointment-booking agent
Speak in a stable branded voice on phone calls without re-testing every month
Pin sonic-3.6-2026-08-27 over the WebSocket API with pcm_mulaw at 8 kHz for telephony. Startup at $49 gives 1.25M credits, roughly 1,600 minutes of speech, and 5 concurrent streams.
A contact-centre operations lead with 100,000 minutes of agent speech a month
Budget TTS for an AI front line replacing hold-queue IVR
At about 800 characters a minute that is 80M credits: Scale's 8M for $299 plus 72M at $38 per million, about $3,000 a month. Eleven v4 Turbo at its $40 rate is about $3,200; Gemini 3.8 Flash TTS on Artificial Analysis's rate about $1,300.
An ed-tech product team serving learners in India
Read lessons aloud in Hindi, Hinglish, Odia and Urdu
Set language or locale per request; 3.6 added Odia and Urdu with instant cloning and improved Hinglish. Opt out of training via the data-controls form before uploading teacher voice samples.
Before you commit
- Setup
- API key from the Cartesia playground; call sonic-3.6 or a dated snapshot over REST or WebSocket. Instant voice cloning needs Pro or above; professional clones need Startup (2 slots) or Scale (4).
- Integrations
- WebSocket streaming with context IDs, flush and cancel; word and phoneme timestamps; PCM and mu-law/A-law output from 8 to 48 kHz; Cartesia's own agent platform and phone numbers.
- Data and privacy
- The privacy policy (June 2024) allows training on uploaded text, audio and outputs unless you opt out via a form, and gives no retention period. SOC 2 Type II company-wide; DPAs, BAAs, SSO and on-premise or VPC deployment on Enterprise only.
- Cost at scale
- About 750 to 800 credits per minute of audio. 100,000 minutes a month is roughly 80M credits, about $3,000 on Scale with overage at $38 per million; Enterprise volume pricing is not published. Agents add $0.06 a minute.
- Limits
- Concurrent TTS streams: 2 on Free, 3 on Pro, 5 on Startup, 15 on Scale. With overages off, requests are blocked when credits run out. The text buffer waits up to 3 seconds by default before generating.
What's good
- Fourth on Artificial Analysis's speech arena, 98 Elo above Sonic 3.5
- Dated snapshots that never change
- Streaming WebSocket API with word and phoneme timestamps
- 44 languages, now including Odia and Urdu
What's not
- Trains on customer content by default; opt-out is a form
- Costs more per character than higher-ranked Eleven v4 Turbo
- No published latency figures
- Free plan has no commercial licence; concurrency starts at 2 streams
Best for
Teams building production voice agents that want stable, versioned voices at a moderate price.
Alternatives
Cartesia Sonic 3.6: common questions
- Is Cartesia Sonic 3.6 worth it?
- Sonic 3.6 closed most of the gap to ElevenLabs in one release and costs less per character than Eleven v4, but it trains on customer content unless you opt out. We rate it 4.4 out of 5.
- How much does Cartesia Sonic 3.6 cost?
- Free plan 20K credits/month (no commercial licence); Pro $5 (100K), Startup $49 (1.25M), Scale $299 (8M); overage $38 to $65 per 1M credits. One credit is one character, about 750 to 800 per minute of audio.
- What are the downsides of Cartesia Sonic 3.6?
- Trains on customer content by default; opt-out is a form; costs more per character than higher-ranked Eleven v4 Turbo; no published latency figures; free plan has no commercial licence; concurrency starts at 2 streams.
- Who is Cartesia Sonic 3.6 best for?
- Teams building production voice agents that want stable, versioned voices at a moderate price.
- How much does Cartesia Sonic 3.6 cost for a team or in production?
- About 750 to 800 credits per minute of audio. 100,000 minutes a month is roughly 80M credits, about $3,000 on Scale with overage at $38 per million; Enterprise volume pricing is not published. Agents add $0.06 a minute.
- Is Cartesia Sonic 3.6 safe for company data?
- The privacy policy (June 2024) allows training on uploaded text, audio and outputs unless you opt out via a form, and gives no retention period. SOC 2 Type II company-wide; DPAs, BAAs, SSO and on-premise or VPC deployment on Enterprise only.
- What are the best Cartesia Sonic 3.6 alternatives?
- The closest alternatives we have reviewed are Eleven v4 and v4 Turbo and Gemini 3.8 Flash TTS.
Answers checked Oct 22, 2026.
Make Cartesia Sonic 3.6? Show this rating on your site
