AssemblyAI vs Cartesia for voice agents: Sonic 3.6, Ink and the speech layer compared

Most "AssemblyAI vs Cartesia" questions start as a TTS question. Which voice sounds better? Which one is faster?

That's the wrong place to start in 2026, because both companies now sell more than a voice. Cartesia sells Sonic 3.6 for text-to-speech, Ink 2 for speech-to-text, and Managed Agents, a hosted agent product billed per minute. AssemblyAI sells Universal-3.6 Pro Realtime for streaming speech-to-text, LLM Gateway for the model, and AssemblyAI-TTS-1 for text-to-speech, as components you run inside LiveKit, Pipecat, or your own orchestrator.

So this comparison goes layer by layer: the voice, the ears, the bill, the concurrency limits, and the languages. Where Cartesia covers something we don't, the post says so. Where a public benchmark has Cartesia ahead, the post says that too.

At a glance

Prices and limits as of October 28, 2026, from each company's own pricing page and docs.

Text-to-speech: AssemblyAI-TTS-1 vs Sonic 3.6

AssemblyAI-TTS-1 is built for one job: conversational voice agents on live calls. It was trained on millions of hours of audio, including real conversational data. It ships with thousands of voices across six languages (English, Spanish, French, German, Italian, and Portuguese), instant voice cloning in the dashboard, and three interfaces: a WebSocket session per conversation, an HTTP endpoint, and an OpenAI-compatible route. Each voice speaks all six languages. The product team describes it as frontier-level quality, faster than other production voice agent TTS models, with unlimited concurrency.

Sonic 3.6 is generally available with 44 languages, according to Cartesia's model docs. Two details matter for voice agent teams:

• Versioning. Cartesia publishes dated snapshots, such as sonic-3.6-2026-08-27, that never change, alongside a sonic-3.6 alias that moves to the newest stable snapshot. AssemblyAI's docs list a single model ID, assemblyai-tts-1.
• Fine control. Cartesia's docs and tooling include pronunciation dictionaries. AssemblyAI TTS has none, and the connection parameter docs say there's no speed, pitch, emotion, style, or stability parameter and no SSML. You steer delivery through the text itself, with punctuation, capitalization, and written-in filler words, as the prompting guide shows.

On telephony the two are even. Both output pcm_mulaw and pcm_alaw at 8 kHz for carrier networks. Cartesia's output format guide and AssemblyAI's connection parameters both name Twilio's 8 kHz μ-law as the target.

How they benchmark. Here is where the two stand on published numbers:

• Entity accuracy (phone numbers, dates, URLs, confirmation codes, drug names): AssemblyAI-TTS-1 93.7% vs Sonic 3.6 82.5%, in our evals from the launch post.
• Median time to first byte: 47 ms vs 143 ms in our evals. That's a single-client lab measurement, one request at a time over each provider's API; the launch post's production time-to-first-audio figure for AssemblyAI-TTS-1 is 90 ms.
• Predicted listener preference (an automated MOS estimate, 1 to 5): 3.85 vs 3.82 in our evals, close to a tie. Inworld TTS-2 Flash scored higher than both at 3.88.
• Word error rate on Coval's TTS leaderboard: 1.7% vs 1.8% as of October 28, 2026, with the leader three-tenths of a point ahead of AssemblyAI-TTS-1. Coval re-measures daily over a rolling 30 days, so the order moves.

Treat those numbers as a starting point, not a verdict. Voice quality depends on your voices, your scripts, and your callers, and latency numbers measure different things (time to first byte versus time to first audible audio). Our TTS latency guide explains the difference, and how to evaluate TTS for voice agents gives you a test you can run on both in an afternoon.

Speech-to-text: Universal-3.6 Pro Realtime vs Ink

A voice agent hears before it speaks, and a misheard order number becomes a wrong answer no voice can fix. So the input side deserves as much attention as the voice.

Universal-3.6 Pro Realtime is AssemblyAI's streaming flagship: 32 languages, built-in turn detection, and $0.45 per hour. Cartesia's current streaming model is Ink 2, released September 17, 2026, covering English, French, Hindi, Japanese, and Spanish. Cartesia lists its earlier model, Ink-Whisper, under "Older Models."

Here's what the AssemblyAI benchmarks page publishes for Cartesia, and it's important to read which Cartesia model each row is.

Our own suites (Cartesia Ink-Whisper). The benchmarks page ran Ink-Whisper, not Ink 2, through its voice-agent and entity suites (runs from August 19 to 30, 2026):

Ink 2 appears in the page's systems table without a current run on those suites, so there's no like-for-like Ink 2 row from our own harness yet.

Third-party tables that include Ink 2. The benchmarks page also reproduces public leaderboards as published, and Ink 2 shows up in three of them:

Read together, those rows say something specific. On clean-clip accuracy and on the Pipecat set, Universal-3.6 Pro Realtime ranks ahead of Ink 2. On Cekura's small real-call set and on the Artificial Analysis index, Ink 2 posts the lower error rate, and Pipecat's median time to final transcript has the two within 8 ms of each other, with Ink 2 ahead. Benchmarks disagree because their audio disagrees, which is the strongest argument for testing on your own calls.

Two differences don't depend on any leaderboard. Universal-3.6 Pro Realtime covers 32 languages against Ink 2's five. And it bills by the hour of audio, while Ink 2 draws 3 credits per second from the same credit pool as Sonic.

For a deeper look at the input side, see how to choose a speech-to-text API for voice agents.

Pricing shape: per character vs credit plans

The two companies bill in different units, so put them in the same spreadsheet before you compare.

AssemblyAI is metered and pay-as-you-go: $15 per 1M characters for TTS and $0.45 per hour for Universal-3.6 Pro Realtime, with no commitment. LLM Gateway bills per model. See the pricing page.

Cartesia sells monthly plans with a credit allotment, per its pricing page:

According to Cartesia's pricing docs, TTS costs about 1 credit per character, Ink 2 costs 3 credits per second of audio, and Managed Agents bill separately at $0.06 per minute, plus $0.014 per minute on a Cartesia-provided number, before LLM usage. Past the allotment, requests either bill as overage or fail, depending on whether overages are enabled for your workspace. The commercial use license starts on Pro.

The practical difference: with credits, TTS and STT draw down one shared monthly pool, so a month's bill depends on your mix and which plan bucket you land in. With per-unit pricing, each layer scales on its own meter.

Concurrency and scaling

This is where the two diverge most.

Cartesia caps concurrent requests per plan: 15 concurrent TTS requests and 60 concurrent Ink 2 requests on Scale, with custom limits on Enterprise. Its pricing FAQ says TTS, STT, and agent concurrency are separate pools. In a voice agent, each reply is a TTS request, so the TTS cap limits how many agents can speak at the same moment.

AssemblyAI counts new sessions per minute and doesn't count sessions that are already open. The TTS limits docs spell it out: a voice agent opens one TTS session per call and keeps it for the whole call, so the limit is how many calls can start in a minute. With a limit of 100, 100 calls can start in one minute and 110 in the next, and all 210 stay connected. The starting limit matches your streaming speech-to-text limit (5 on a free plan, 100 or more on a paid plan) and grows with use on paid plans.

That's what unlimited concurrency means in practice: capacity that grows with your call volume and never requires negotiation. Sessions still last at most an hour and close after two minutes idle without a keepalive, and the LiveKit plugin and Pipecat service handle both.

Languages and versioning

Languages. Sonic 3.6 speaks 44 languages, including Hindi, Japanese, Korean, Arabic, and many Indian languages. AssemblyAI-TTS-1 speaks six. On the input side it flips: Universal-3.6 Pro Realtime covers 32 languages, and Ink 2 covers five. If your agent speaks, say, Korean, Cartesia's TTS covers it and AssemblyAI's doesn't, and you'd need a speech-to-text model that covers Korean too.

Versioning. Cartesia's dated Sonic snapshots never change, which lets you pin a voice and re-run evals before you move. AssemblyAI exposes assemblyai-tts-1 as one model ID. If your compliance process requires pinning an exact model build, factor that in.

Data zones. AssemblyAI TTS runs in US and EU data zones (tts.us.assemblyai.com and tts.eu.assemblyai.com), as the endpoints docs describe.

Which to choose

The bigger takeaway is that "AssemblyAI vs Cartesia" isn't one comparison anymore. It's three, and you don't have to pick the same answer for each. Plenty of teams will run one vendor's ears with another's voice. Benchmark each layer on your own calls, then decide whether one bill and one key are worth more to you than the best score on any single layer. For the full component view, see the voice AI stack for building agents.

Frequently Asked Questions (FAQs)

What is Cartesia Sonic?

Sonic is Cartesia's text-to-speech model family. The current version, Sonic 3.6, is generally available, speaks 44 languages, and comes with dated snapshots that never change plus a sonic-3.6 alias that tracks the newest stable snapshot. It's billed at about 1 credit per character from Cartesia's monthly credit plans.

How much does Cartesia cost?

Cartesia sells monthly plans: Free ($0, 20K credits), Pro ($5, 100K), Startup ($49, 1.25M), Scale ($299, 8M), and custom Enterprise, as of October 28, 2026. TTS costs about 1 credit per character and Ink 2 costs 3 credits per second of audio. Managed Agents bill separately at $0.06 per minute plus LLM usage.

Is AssemblyAI TTS better than Cartesia Sonic for voice agents?

It depends on your languages and volume. AssemblyAI-TTS-1 covers six languages, counts new sessions per minute instead of concurrent requests, and costs $15 per 1M characters. Sonic 3.6 covers 44 languages, offers pinned snapshots and pronunciation dictionaries, and caps concurrent requests per plan. Test both on your own scripts.

How does Universal-3.6 Pro Realtime compare to Cartesia Ink 2?

Universal-3.6 Pro Realtime covers 32 languages to Ink 2's five. On tables reproduced on AssemblyAI's /benchmarks page, Universal-3.6 Pro Realtime ranks ahead on Cekura's clean clips and the Pipecat set, while Ink 2 posts lower error on Cekura's real-call set and the Artificial Analysis index. Run both on your own audio.

How does Cartesia compare to ElevenLabs?

Both sell TTS, STT, and hosted agents, but they bill differently. Cartesia uses monthly credit plans at about 1 credit per character, and Sonic 3.6 covers 44 languages. ElevenLabs lists v4 Turbo at $0.04 per 1K characters with 90+ languages. Both cap concurrency per plan, so check how each counts it.

Can I use Cartesia TTS with AssemblyAI speech-to-text?

Yes. In LiveKit or Pipecat, speech-to-text and text-to-speech are separate plugins, so you can pair Universal-3.6 Pro Realtime with Sonic, or Ink with AssemblyAI-TTS-1. You'll manage two API keys and two bills. Running all layers on AssemblyAI keeps it to one key.