Cartesia Sonic Explained: Models, Latency, Pricing

Cartesia Sonic is Cartesia's streaming text to speech model family. The current stable model is Sonic 3.6, released August 27, 2026, with 44 languages. Cartesia claims first audio in under 90 milliseconds, while one third-party test measured Sonic 3.5 at a 128 ms median. Older Sonic models stop working after October 20, 2026.

What is Cartesia Sonic?

Sonic is a family of streaming text to speech models that Cartesia sells through its API. AWS describes Sonic 3 as "Cartesia's latest state space model (SSM) for streaming text-to-speech (TTS)," and it lists 42 languages for that version. Cartesia also offers a self-hosted Sonic 3.6 release, published on September 17, 2026, according to its changelog.

Because Sonic is an API product, its pricing and features are aimed at developers and voice agents. If you want a studio for finished narration, the app-based options matter more. Our PlayHT alternatives guide compares both kinds of tool.

Sonic model versions and what is retiring

Cartesia's model documentation says the alias sonic-3.6 updates automatically, while dated snapshots "will never change." The table below uses Cartesia's release and sunset dates.

| Model | Released | Languages | Status | |---|---|---|---| | sonic-3.6 (stable snapshot 2026-08-27) | 2026-08-27 | 44, with Odia and Urdu new | Current stable model | | sonic-3.5-2026-05-04 | 2026-05-04 | 42 | Stable | | sonic-3-2026-01-12 | 2026-01-12 | 42 | Stable | | sonic-3-2025-10-27 | 2025-10-27 | 42 | Stops working after 2026-10-20 | | sonic-2 snapshots | April to June 2025 | 8 | Stop working after 2026-10-20 | | sonic-turbo snapshots | June 2025 | 9 | Stop working after 2026-10-20 |

The older models page confirms that sonic-2, sonic-turbo and the sonic-3-2025-10-27 snapshot stop working after October 20, 2026. If your code uses any of them, move it to sonic-3.6 now. The sonic-preview model is labeled beta.

How fast is Sonic? Claimed versus measured

There is no single latency number for Sonic. The figures below come from different sources, and only one is a measurement by a third party, covering Sonic 3.5 rather than 3.6.

| Source | Figure | Model | Type | |---|---|---|---| | Cartesia Sonic page | First audio in under 90 ms | Sonic (undated) | Vendor claim | | Sonic 3.5 documentation | Sub-90 ms latency | Sonic 3.5 | Vendor claim | | Sonic 2 launch, March 2025 | 90 ms for the full model, 40 ms for turbo | Sonic 2 | Vendor claim | | Sonic 3 changelog, October 2025 | About 190 ms median | Sonic 3 | Vendor changelog | | Vapi Humanness Index | 128 ms median time to first audio over 50 live trials, June 2026 | Sonic 3.5 | Measured, includes network round trip, one test setup |

The Sonic 3.6 documentation gives no latency figure, and Vapi did not measure Sonic 3.6. Use the measured figure as a single data point for one model, not a guarantee for every network.

Sonic features

Sonic's controls are set in the request configuration. The main options are:

• Speed: generation_config.speed runs from 0.6 to 1.5, with a default of 1.0. Cartesia notes that this is guidance, not a strict multiplier.
• Volume: generation_config.volume runs from 0.5 to 2.0.
• Emotion (beta): generation_config.emotion offers 58 values, including neutral, calm, angry, content, sad and scared. It works in English only.
• Laughter: insert [laughter] into the transcript.
• Inline tags: SSML-style tags such as <emotion value="angry"/> are supported.
• Sonic 3.6 pacing: the model paces speech from emotional context without SSML tags. Disfluencies such as "uhm" and "hmm" in the transcript give a slower, thinking pace.
• Pronunciation dictionaries: on 3.6, they ignore case by default.

Cartesia's volume, speed and emotion page lists these settings in full. Speed is not available for Professional Voice Clones at request time.

Voice cloning on Cartesia

Cartesia offers two kinds of voice clone, and the plan you choose decides which ones you can use. Instant cloning is on Pro and above. Professional clones are on Startup and Scale, with 2 slots on Startup and 4 on Scale. Cartesia's own page says a voice can be cloned from 10 seconds of audio, which is a vendor claim.

Cartesia says instant clones keep more accent on Sonic 3.6, and that professional voice clones made on older models work on 3.6. Training a Professional Voice Clone is free, and generating audio from one costs 1 credit per character.

Cartesia pricing: credits, plans and cost per million characters

Cartesia's pricing page bills in credits. One credit equals one character, and one minute of audio needs roughly 750 to 800 credits.

| Plan | Monthly price | Credits per month | Concurrency | Cloning | Cost per 1M characters (our arithmetic) | |---|---|---|---|---|---| | Free | $0 | 20,000 | 2 | None | Not applicable | | Pro | $5 | 100,000 | 3 | Instant | $50.00 | | Startup | $49 | 1,250,000 | 5 | Instant plus 2 Professional slots | $39.20 | | Scale | $299 | 8,000,000 | 15 | Instant plus 4 Professional slots | About $37.38 | | Enterprise | Custom | Custom | Custom | Custom | Custom |

Using 750 to 800 credits per minute, our arithmetic gives about 25 to 27 minutes on Free, 125 to 133 minutes on Pro, 1,563 to 1,667 minutes on Startup and 10,000 to 10,667 minutes on Scale. Overage is billed at $65, $45 and $38 per million credits on Pro, Startup and Scale, but only if overage is enabled. Without it, requests stop when credits run out.

For voice agents, Cartesia lists $0.06 per call minute on all plans, and telephony at $0.014 per minute with a Cartesia phone number. Its LLM usage offer was free until October 1, 2026. That date has passed, and we did not verify how LLM usage is billed now.

How Sonic ranks in blind listening tests

The Artificial Analysis TTS leaderboard is an undated, blind-preference arena. Its newest model is from October 2026. Sonic 3.6 ranks fourth with 1285 Elo, behind Eleven v4 Turbo at 1330, Eleven v4 at 1329 and one other model. Sonic 3.5 ranks 17th with 1193 Elo.

Cartesia's own page claims 1284 Elo and calls the model number one on the Speech Arena. We use the independent figure and do not repeat the number one claim. Elo measures listening preference, not price or latency.

The Vapi Humanness Index gives Sonic 3.5 a score of 70 and a ranking of about 12th, from 1,094 blind votes. Listeners compared two voices reading a customer support prompt at phone quality.

Who Sonic is for, and who should look elsewhere

Sonic fits teams that need:

• Streaming audio for a live call or a voice agent.
• Per-character API pricing that they can meter in their own code.
• Many languages, with 44 in Sonic 3.6.

Look elsewhere if you need:

• Finished narration, dialogue or voiceover produced without writing code.
• Emotion controls in a language other than English, since emotion is an English-only beta.
• A clear, fixed monthly bill for a small team, where a studio plan may be easier to manage.

Where OndaLabs fits

Cartesia sells an API for developers and voice agents, metered per character. OndaLabs is a voice studio that you use from the app, with text to speech, two-speaker dialogue, speech to text, custom voices, image generation and music generation. We do not compare speed, and we do not present OndaLabs as a real-time agent tool. If you need streaming to a live call, Cartesia is built for that. If you are producing finished audio such as narration, dialogue or voiceover, a studio like OndaLabs is another route.

Both products use per-character credits, which allows a rough price comparison. OndaLabs Starter costs $9 for 170,000 credits, or about $52.94 per million credits. Cartesia Pro costs $5 for 100,000 characters, or $50.00 per million. These are different products, tiers and quality levels, so no winner is implied.

OndaLabs controls delivery by describing it in plain words and with supported expressive tags in the script. That is a similar idea to Sonic's emotion and laughter controls, not the same feature. You can try text to speech in the app, build a conversation with dialogue, or compare plans and credits.

Frequently asked questions

What is the latest Cartesia Sonic model?

Sonic 3.6 is the current stable model. The sonic-3.6 alias updates automatically, and the dated snapshot sonic-3.6-2026-08-27 does not change. It supports 44 languages, and Odia and Urdu are new in this version.

How fast is Cartesia Sonic?

Cartesia claims first audio in under 90 ms. A third-party test measured Sonic 3.5 at a 128 ms median time to first audio over 50 live trials, including network time. The test covered one setup, and it did not measure Sonic 3.6.

Does Cartesia support voice cloning, and how much audio does it need?

Yes, on paid plans. Instant cloning is on Pro and above, and Professional clones are on Startup and Scale. Cartesia says a voice can be cloned from 10 seconds of audio, which is a vendor claim. Professional clone training is free, and generating from one costs 1 credit per character.

How much does Cartesia cost per minute?

Cartesia bills 1 credit per character, and one minute of audio needs roughly 750 to 800 credits. Pro costs $5 a month for 100,000 credits, which is about 125 to 133 minutes by that estimate. Startup and Scale give a lower cost per million characters.

Is Cartesia better than ElevenLabs for voice agents?

The independent Elo scores rank Eleven v4 Turbo above Sonic 3.6, but Elo measures blind listening preference, not agent fit. Cartesia offers a voice agent price of $0.06 per call minute. Test both on your own calls before you choose.

Which languages does Sonic support?

Sonic 3.6 supports 44 languages and 61 locales, according to Cartesia's changelog. Emotion controls are English-only. Check the language list in the current documentation before you build for a new market.