Voxtral 3 TTS vs Cartesia Sonic 3.6: 72 Elo for 3.8 Times the Price

Cartesia Sonic 3.6 wins the quality argument and Voxtral 3 T​TS wins the invoice, and the size of both margins is the whole comparison. On Artificial Analysis's Provider Voice leaderboard on October 21, 2026, Sonic 3.6 sits at 1,289 Elo (#4 of 98 rated systems) against 1,217 for Voxtral 3 T​TS (#13). The gap is 72 points, comfortably outside the two models' confidence intervals. The price gap runs the other way and is wider: $49 per million characters for Sonic 3.6 against $13 for Voxtral 3 T​TS — 3.8 times as much for a model that Mistral shipped one day earlier than this reading of the board. Neither number is in dispute. What is in dispute is how much a 72-point blind-listening margin is worth in production, and that depends entirely on what the voice is doing. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

First, what the boards actually measured

Both figures come from the same source, which matters more than usual in a category where every vendor publishes its own preference study. Artificial Analysis runs two boards, and the two models appear on both.

• Provider Voice — each model is judged with its own native voices. Sonic 3.6: 1,289 Elo, #4, 95% interval 1,275–1,303, 2,732 comparisons. Voxtral 3 T​TS: 1,217 Elo, #13, interval 1,202–1,232, 1,619 comparisons.

• Controlled Voice — every model is cloned onto the same eight reference voices, so listeners compare the engine rather than the voice casting. Sonic 3.6: 1,151 Elo, #5, interval 1,138–1,164. Voxtral 3 T​TS: 1,066 Elo, #14, interval 1,049–1,083.

• Released — Sonic 3.6 in August 2026; Voxtral 3 T​TS on October 20, 2026.

• List price — $49.00 per 1M characters versus $13.00 per 1M characters, both as tracked by Artificial Analysis and matching each vendor's public pricing.

• Throughput — 145 characters per second median for Sonic 3.6 versus 112 for Voxtral 3 T​TS, measured by Artificial Analysis.

• Weights — closed for both. Neither can be self-hosted under a permissive licence.

Two details about that table are easy to skip and shouldn't be. The Controlled Voice result is the more honest of the two, because it removes the advantage of having a good voice library, and there the gap widens from 72 points to 85. Sonic 3.6 is not winning on voice casting; it is winning on synthesis. The second detail is the sample count. Voxtral 3 T​TS has 1,619 comparisons on the provider board, roughly 40% fewer than Sonic 3.6, and its interval is correspondingly wider. A day-two score is a real score, but it will move.

Where 72 Elo comes from, feature by feature

Elo is a preference sum, not a diagnosis, so it is worth asking which product decisions produced it. The two models were built for different jobs and the specs show it.

Cartesia built a streaming engine. Sonic 3.6's headline claim is first audio in under 90 milliseconds, from a vendor page that leads with speed. Its controls are the ones a live agent needs: inline cues such as laughter, volume and speed modulation, custom pronunciation dictionaries, emotion inferred from context. It is a model designed to be called mid-conversation and to start talking almost immediately.

Mistral built a studio voice with an enterprise wrapper. Voxtral 3 T​TS offers 35+ languages and 50+ dialects, voice presets in 18 languages, prompt-level control over emotion, pacing and emphasis, phoneme-level pronunciation, and a vocabulary library. Its cloning story is "Pro Voice": 30 minutes of audio to reproduce a speaker's character, emotional range and style across languages. That is an adaptation pipeline, not an instant clone.

The practical consequence is that the two models are not competing for the same slot. A 30-minute voice-adaptation process is a brand-voice operation; a 10-second instant clone is a user-facing feature. If your product lets a customer upload a clip and hear themselves in a generated voice, Sonic 3.6 does that and Voxtral 3 T​TS, as documented, does not.

The latency asymmetry compounds it. Cartesia publishes a latency number and Mistral does not. Voxtral 3 T​TS has no published time-to-first-audio anywhere in the launch material; the closest figure is the sub-500ms claim on the separate realtime transcription model. A missing number is not a bad number, but in a buy decision it is a risk you carry, and the vendor that publishes its number is the safer default for anything interactive.

The cost arithmetic, at the scale where it bites

Per-character pricing looks abstract until it is multiplied by volume. At roughly 900 characters per spoken minute, one hour of continuous speech is about 54,000 characters.

• One hour of audio — Sonic 3.6: about $2.65. Voxtral 3 T​TS: about $0.70.

• A thousand hours a month — Sonic 3.6: about $2,650. Voxtral 3 T​TS: about $700. A difference of roughly $23,400 a year.

• Ten thousand hours a month — Sonic 3.6: about $26,500. Voxtral 3 T​TS: about $7,000. A difference of roughly $234,000 a year.

At the lower end of that range the gap is a line item someone argues about. At the upper end it is a headcount. That is the entire case for Voxtral 3 T​TS, and it is a strong one for any workload where the voice is long-form and pre-rendered: audiobooks, e-learning modules, catalogue narration, dubbing pipelines. In those jobs the 72 Elo shows up as a slightly less natural read on a small fraction of sentences, and the listener is not comparing it to anything.

Where the arithmetic changes shape is when the two models are doing different jobs in one product. A support line needs Sonic 3.6's latency and its sub-second start; the same company's marketing narration does not. Paying $49 for the second job is $49 spent on a capability the job cannot use — and the cleanest way to keep both without two procurement processes is a routing layer that treats the model as a parameter. That is the argument for OrcaRouter's setup rather than for either vendor: one OpenAI-compatible API across 200+ models, provider list price passed through at 0% markup, and a routing DSL that can send a batch narration job and a realtime call to different backends on the same key. Neither Sonic 3.6 nor Voxtral 3 T​TS is in OrcaRouter's catalogue as of this writing, so today both are reached through their vendors' own APIs and several third-party platforms — the point is the shape of the integration when they are.

Three objections that decide it in practice

The first is language. Cartesia advertises 44 languages from one model. Mistral advertises 35+ languages and 50+ dialects, with presets in 18. On paper Mistral covers more ground; in practice the dialect count is a promise about coverage breadth, not a promise that every dialect has been evaluated. If your roadmap includes a market neither vendor has a demo for, ask for a sample rather than reading the count. The one place Mistral is unambiguously stronger is European coverage: it claims all 24 official EU languages for its transcription models and 35+ for synthesis, and Mistral is a French company selling into EU procurement with data-residency options in EU, US and global endpoints. Sonic 3.6 has no equivalent sovereign-deployment story.

The second is pronunciation control in production. This is where the Elo gap misleads most. A model that reads "Dr. Nguyen" correctly without a dictionary can rank below one that sounds more natural on generic prose. Voxtral 3 T​TS ships a vocabulary library and phoneme control; Sonic 3.6 ships custom pronunciation dictionaries and inline delivery cues. Both have the feature. Mistral's is aimed at names and brand terms in enterprise content; Cartesia's is aimed at the same problem inside a streaming call. Test your own lexicon — a list of product names, surnames and acronyms read aloud — before either score gets a vote.

The third is what happens when a vendor has a bad afternoon. A single-endpoint voice pipeline fails closed: the endpoint degrades and the product goes quiet. Sonic 3.6 and Voxtral 3 T​TS sit on different infrastructure with different failure modes, and the fallback path between them is worth designing before it is needed. Automatic failover across providers is the specific thing a router buys you here, and it is more valuable in speech than in text because a dropped token is invisible and a dropped syllable is not.

So which one

Pick Cartesia Sonic 3.6 when the voice is in a conversation: phone agents, in-app assistants, game NPCs, anything where the first syllable arrives before the user has finished reading the spinner. You are paying $49 a million characters for a model that starts in under 90 milliseconds and scores 1,289 on the board, and the price is defensible against every other model in that latency class.

Pick Voxtral 3 T​TS when the voice is content: narration, dubbing, brand voice across a catalogue, anything rendered ahead of time and played back. You are paying $13 for a model that is 72 Elo behind the leader and 3.8 times cheaper, with EU data residency and a 30-minute voice-adaptation pipeline that suits a single corporate voice better than a per-user clone.

Pick neither exclusively if you ship both kinds of audio. The mistake to avoid is choosing once and then paying the premium model's rate for work the cheap model can do, or absorbing the cheap model's latency in a place where it is fatal. Two endpoints behind one key is the configuration that stops the decision from being permanent.

The open question is what Mistral's missing latency number turns out to be. If Voxtral 3 T​TS streams with a first-audio figure anywhere near Cartesia's, the $13 price becomes a much harder fact to argue with and the 72 Elo becomes the only thing Sonic 3.6 still has. Mistral has not published it, no third party has measured it, and until one of those changes, the streaming slot belongs to Cartesia by default.