Cartesia Sonic 3 and 3.5: Specs, Real Latency, Pricing, and Where to Use It (2026)
Last updated August 7, 2026
Cartesia Sonic is SSM-architecture streaming TTS: Sonic 3 (Oct 2025, 42 languages, emotion and laughter, ~190 ms median) and Sonic 3.5 (GA May 2026: natural pacing, alphanumeric read-out, heteronyms, multilingual step-change). Real-world latency measures ~166-190 ms; sub-90 ms is a vendor model-latency claim. Ten-second cloning, HIPAA/SOC 2, commercial license from $5/month. Both generations run in invideo's Audio tab.
Updated August 2026
Cartesia Sonic is a streaming text-to-speech family built for real-time voice — and it's architecturally unusual: instead of a transformer, Sonic runs on Cartesia's state space model (SSM) design, which is what lets it start speaking in a fraction of a second. As of August 2026 the family has two live generations, Sonic 3 (October 2025) and Sonic 3.5 (GA May 2026), covering 42 languages, ten-second voice cloning, and inline nonverbal tags like [laughter]. The honest headline: Sonic is one of the fastest production TTS engines you can stream, with real-world measured latency around 166–190 ms — the sub-90 ms figure you'll see quoted is a vendor model-latency claim under ideal conditions, and the two numbers should never be conflated.
What is Cartesia Sonic?
Sonic is low-latency, expressive TTS aimed first at voice agents — anything where a pause reads as a failure — and second at high-volume voiceover. The SSM architecture processes audio as a continuous stream rather than in chunks, the design bet behind its speed. Vapi's model brief put the release in context: "Sonic 3 landed in October 2025 as a major step for Cartesia's state space model architecture."
Sonic 3 vs Sonic 3.5: what changed?
| Variant | Released | Documented additions |
|---|---|---|
Sonic 3 (sonic-3-2025-10-27) |
Oct 27, 2025 | 42 languages; volume, speed, and emotion controls; real-time conversation with emotion and laughter; ~190 ms median latency |
Sonic 3.5 (sonic-3.5-2026-05-04) |
Preview Apr 2026, GA May 2026 | More natural pacing and emotional expression; cleaner audio; alphanumeric read-out; heteronym handling; multilingual step-change |
Per Cartesia's changelog, Sonic 3 shipped October 27, 2025 with "42 languages; volume, speed, and emotion controls… Real-time conversation with emotion and laughter; ~190ms median latency." Sonic 3.5 previewed on sonic-3-latest in April 2026 and went GA in May 2026 as the current stable snapshot.
The 3.5 upgrades are unusually practical. It reads alphanumerics correctly — confirmation codes, order numbers, phone numbers, email addresses — which is precisely where voice agents used to embarrass themselves. It resolves English heteronyms ("read," "bass," "bow") from context. And Cartesia describes a "step-change multilingual performance, particularly Hebrew, Japanese, Spanish, Hindi, German, Korean, and French." Professional Voice Clones have run on 3.5 since June 2026, per the models documentation.
How fast is Cartesia Sonic really?
Three numbers circulate, and they measure different things:
| Figure | What it is | Source |
|---|---|---|
| Sub-90 ms | Model time-to-first-audio under ideal conditions | Cartesia's marketing page — vendor claim |
| ~190 ms median | Real-time conversation latency | Cartesia's own changelog (Oct 2025) |
| 166 ms median | Streaming latency including network round-trip | Vapi Humanness Index benchmark, June 2026 — third-party measurement |
Plan around the real-world band: ~166–190 ms to first audio. That is still fast enough that a voice agent answers inside natural conversational turn-taking — just don't put "sub-90 ms" in a latency budget as if it included your network.
Cartesia Sonic's spec sheet
From the official product page and docs, August 2026:
| Spec | Detail |
|---|---|
| Languages | 42 |
| Voice cloning | Instant cloning from 10 seconds of audio (Pro plan); professional cloning at Startup+ |
| Nonverbal tags | Inline in text, e.g. [laughter] |
| Pronunciation | Custom pronunciation dictionaries |
| Compliance | HIPAA, SOC 2 Type 2, GDPR, PCI |
The compliance row is the sleeper here: HIPAA and SOC 2 Type 2 attestations are rare among TTS vendors, which makes Sonic one of the few real options for healthcare and regulated-industry voice work.
How much does Cartesia Sonic cost?
Official pricing, August 2026:
| Plan | Price | Includes |
|---|---|---|
| Free | $0 | 20K credits/month |
| Pro | $5/month | 100K credits, commercial license, instant cloning |
| Startup | $49/month | 1.25M credits, professional cloning |
| Scale | $299/month | 8M credits |
| Enterprise | Custom | Voice-agent calls at $0.06/min; telephony $0.014/min |
The $5/month commercial entry point is among the lowest in production TTS — but note that free-tier usage does not carry the commercial license; that starts at Pro.
How good does Sonic actually sound?
Cartesia claims a #1 ranking for naturalness on the Artificial Analysis Speech Arena — a vendor claim, and leaderboards move monthly, so treat any "#1" as time-stamped. The independent picture is more nuanced: Vapi's blind-voted Humanness Index (June 2026 snapshot) scores Sonic 3.5 at 71, ranked #13, up from Sonic 3's 66 — a genuinely good score on a crowded board, not an uncontested lead. Sonic's case rests on the speed-quality-price combination, not on winning every blind listening test.
On adoption, Cartesia's own site names ServiceNow and Quora as enterprise customers (vendor-claimed), and Sonic is offered as a TTS option inside Vapi's voice-agent platform.
What can you make with Cartesia Sonic?
- Real-time voice agents — the family's home turf: sub-200 ms measured streaming, alphanumeric accuracy for codes and order numbers, and pronunciation dictionaries for product names.
- Voiceover at scale — 42 languages and credit-based pricing suit batch narration through an AI voice generator workflow.
- Script-to-speech production — 3.5's pacing and heteronym handling make it a clean engine behind an AI text to speech pass, where a mispronounced "read" would otherwise mean a retake.
Cartesia Sonic, quickly answered
What is the difference between Sonic 3 and Sonic 3.5?
Sonic 3.5 (GA May 2026) adds more natural pacing, correct alphanumeric read-out (confirmation codes, phone numbers), context-aware heteronym pronunciation, cleaner audio, and a documented multilingual step-change across Hebrew, Japanese, Spanish, Hindi, German, Korean, and French. Sonic 3 (October 2025) established the 42-language, emotion-and-laughter baseline.
Is Cartesia Sonic really sub-90 ms?
Sub-90 ms is Cartesia's model-latency claim under ideal conditions. Measured real-world figures are ~166 ms median including network (Vapi, June 2026) and ~190 ms median per Cartesia's own changelog. Budget for the measured band.
How much audio does Sonic voice cloning need?
Instant cloning works from about 10 seconds of audio and is included from the Pro plan ($5/month); higher-fidelity professional cloning requires Startup ($49/month) or above.
Can I use Cartesia Sonic commercially?
Yes, from the Pro plan at $5/month, which includes the commercial license. The free tier (20K credits/month) does not.
Is Cartesia Sonic HIPAA compliant?
Cartesia lists HIPAA, SOC 2 Type 2, GDPR, and PCI compliance on its official site as of August 2026 — one of the few TTS vendors covering healthcare-grade requirements.
Where can you use Cartesia Sonic?
Both generations run inside the invideo agent — the Audio tab's picker carries Cartesia Sonic 3.5 and Sonic 3 — so the engine that powers phone agents can voice your video scripts with no Cartesia account, credit math, or streaming code involved. The invideo AI models index shows the rest of the 200-plus model roster that can run in the same project.
Version history: Sonic 3 stable snapshot (sonic-3-2025-10-27) — October 27, 2025 → Sonic 3.5 preview on sonic-3-latest — April 2026 → Sonic 3.5 GA (sonic-3.5-2026-05-04) — May 2026 → Professional Voice Clones on 3.5 — June 2026. Current as of August 2026 per Cartesia's official documentation, changelog, and pricing pages; third-party latency figures from Vapi's June 2026 benchmark.