Cartesia Ships Sonic-3.6: Streaming TTS Now Leads Both Speech Arenas

Cartesia Ships Sonic-3.6: Streaming TTS Now Leads Both Speech Arenas

On August 17, 2026, Cartesia released Sonic-3.6 in beta — just three months after Sonic-3.5. The headline claim is verifiable: at launch, Sonic-3.6 sat at #1 on both Artificial Analysis speech leaderboards — 1,283 Elo on the Provider Voice board and 1,123 on the Controlled Voice board (August 18, 2026 snapshot). The second number matters more. That board clones every model onto the same eight reference voices, isolating the synthesis engine from the voice catalog. If you top that ranking, the model itself got better — not just the voice library.

Cartesia is not open-sourcing anything. Sonic-3.6 is a hosted API, available on sonic-preview. No weights, no Hugging Face repo. You rent it. Full details on Cartesia's Sonic page.

What Changed From 3.5

The version gap is only three months, but the improvements land in specific places:

• Naturalness across languages. Cartesia describes a "step change in naturalness" across its supported languages.
• Locale controls. New locale and normalization fields separate the voice's accent from reading conventions for dates, numbers, and currency — though English regional read-outs still inherit the same conventions for now.
• English heteronyms. Words like "read," "bass," and "bow" now resolve correctly in context, per the model docs.
• Preview track. Sonic-3.6 is served on the preview track (sonic-preview) while Sonic 3.5 remains the stable default.

The architecture remains state space models, not transformers. Cartesia frames this as the reason for the latency advantage: sub-90ms time-to-first-audio on their benchmark.

What the Leaderboard Actually Measures

Artificial Analysis runs two separate TTS leaderboards. Understanding the difference matters if you're evaluating providers:

Provider Voice tests each model with the vendor's own best voices. This measures the total experience — synthesis quality plus voice catalog. At launch, Sonic-3.6 led here with 1,283 Elo, ahead of Sonic-3.5 and ElevenLabs Eleven v3.

Controlled Voice is the apples-to-apples comparison. Every model runs on the same eight reference voices, so voice quality is constant. The only variable is the synthesis engine. Sonic-3.6 led this board too at 1,123 Elo (August 18, 2026 snapshot), with Sonic-3.5 second and ElevenLabs third.

That second result is the one that should worry competitors. It means Cartesia's synthesis engine — the part that turns text into audio waveforms — genuinely improved. Not "we added better voices."

How to Read the Two Elo Numbers

Provider Voice is a bundle score: engine plus catalog. A lead there can come from better voices alone. Controlled Voice strips the catalog away, so a lead there is the stronger claim — the engine itself got better. One caution that applies to both: live leaderboards drift as votes accumulate, so treat 1,283 and 1,123 as a dated snapshot (18 August 2026), the way you would treat any benchmark snapshot that moves with its harness, not as permanent ratings.

Getting Started

The API surface is straightforward. Here's a minimal TypeScript example:

import { CartesiaClient } from "@cartesia/cartesia-js";

const client = new CartesiaClient({
  apiKey: process.env.CARTESIA_API_KEY,
});

const audio = await client.tts.bytes({
  modelId: "sonic-preview",
  transcript: "Hello from Sonic 3.6. This is streaming text-to-speech.",
  voice: {
    mode: "id",
    id: "820a4f59-ad9b-449c-a45b-480eeadb6773",
  },
  language: "en",
  outputFormat: { container: "raw", encoding: "pcm_f32le", sampleRate: 24000 },
});

// Stream the audio chunks
for await (const chunk of audio) {
  process.stdout.write(chunk);
}

Switching from 3.5? Change the modelId to sonic-preview and keep everything else identical.

For inline expressions, drop tags directly into the transcript:

const transcript = `
That's hilarious! [laughter]
But seriously, the order number is 
<spell>A-T-7-2-9-1-4</spell> and your 
confirmation code is <spell>B-X-0-3</spell>.
`;

Cartesia's prompting guide covers <spell>-style tags for spelling out codes inline — you just get better-sounding output now.

The Pricing Question

At $49 per million characters, Sonic-3.6 sits in a specific tier. Half the price of ElevenLabs Eleven v3 at $100/M, but well above Speechify Simba 3.2 at $10/M (which scores 1,240 Elo on the Provider Voice board — roughly 43 points behind the launch snapshot).

For production workloads, Cartesia sells credits through tiers — see the pricing page. The Pro plan starts at $5/month. Scale at $299/month includes roughly 10,667 TTS minutes and 15 concurrent requests. Voice agents through their Line product bill separately at $0.06 per minute.

The real cost comparison depends on your use case. For high-volume contact centers, the latency advantage translates directly to shorter calls and lower per-interaction cost. For content creation, the quality gap between 1,283 and 1,240 Elo may not justify a 5x price difference.

What This Means for Voice Agent Builders

If you're building voice agents, Sonic-3.6's Controlled Voice leadership is the number to watch. That leaderboard isolates synthesis quality from voice selection, which means the engine itself handles agent transcripts better — confirmation codes, heteronyms, inline expressions — without preprocessing.

Cartesia's Line product already bundles Sonic TTS with Ink-2 STT (speech-to-text). The combined pipeline runs at sub-90ms TTS latency and 100ms transcript latency, both vendor-stated. For real-time voice agents where every millisecond of round-trip time matters, that stack is competitive.

The catch: Sonic-3.6 is beta. Cartesia's docs still list Sonic 3.5 as the stable, production-ready model. The sonic-preview model ID tracks the beta, so if you're shipping to production today, pin to sonic-3.5. If you're evaluating for a future deployment, Sonic-3.6 is worth testing now.

Beta-Pin Checklist

Before switching a production voice stack to the beta, re-run the article's own hard cases: heteronyms (read/bass/bow), dates, and currency — the exact inputs where a synthesis upgrade either proves itself or embarrasses you. And remember the tradeoff the launch implies: with a rented API you get the new engine the day it ships, but you also inherit its beta status — the rented-API tradeoff cuts both ways.

Related

• Needle 2: A 14MB Model That Calls Tools on Any Device — Edge-native tool calling that runs entirely offline.
• When AI Agents Start a Turf War — What happens when multiple Claude agents fight over the same codebase.

Sources

• Artificial Analysis: Text to Speech Leaderboard (Provider Voice)
• Artificial Analysis: Controlled Voice Leaderboard
• MarkTechPost: Cartesia ships Sonic-3.6: a streaming TTS model that now leads both Artificial Analysis speech arenas
• Cartesia: Introducing Sonic-3.6 and Ink-2
• Cartesia: Introducing Ink-2: the #1-ranked STT built for voice agents
• Cartesia Docs: Sonic models and language support
• Cartesia Docs: Locale-aware normalization in Sonic
• Cartesia: Cartesia pricing