Voxtral 3 TTS vs ElevenLabs: Six Times Cheaper, One Hundred Points Behind

Mistral priced Voxtral 3 T​TS at $13 per million characters and shipped it into a market where the model everyone benchmarks against, ElevenLabs Eleven v4, costs $80 and scores 1,328 Elo against Voxtral's 1,217. That is 111 points of blind-listener preference for 6.2 times the money, and on the face of it the cheap model is not close enough to matter. It matters anyway, for a reason the scoreboard does not show: ElevenLabs is not only selling synthesis. It is selling a voice platform — a library, dubbing, agents, professional cloning — and most of the teams paying $80 a million characters are paying for the platform as much as for the voice. Voxtral 3 T​TS, released October 20, 2026, arrives inside an agent SDK and an EU-sovereign deployment story instead. The comparison that decides a purchase is therefore not "which sounds better", because that one is settled, but "which of those two bundles is the one you actually need". 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.

The scoreboard, before anything else

• Provider Voice Elo — ElevenLabs Eleven v4: 1,328 (#2 of 98). Voxtral 3 T​TS: 1,217 (#13). Both per Artificial Analysis, read October 21, 2026.

• Controlled Voice Elo — Eleven v4: 1,157 (#4, tied with Eleven v4 Turbo). Voxtral 3 T​TS: 1,066 (#14).

• List price per 1M characters — $80.00 vs $13.00.

• Languages — ElevenLabs: 74 across all plans. Voxtral 3 T​TS: 35+ languages and 50+ dialects, with presets in 18.

• Instant cloning — ElevenLabs: from roughly 30 seconds of audio on paid plans. Voxtral 3 T​TS: no instant clone published; "Pro Voice" adaptation needs 30 minutes of audio.

• Weights — closed for both.

One line in that list deserves a second look: ElevenLabs' own cheaper flagship. Eleven v4 Turbo scores 1,333 — five points above Eleven v4, inside the noise — at $40 per million characters, exactly half the price. If the choice is purely "best voice ElevenLabs sells", Turbo is the sensible buy and it cuts the gap to Voxtral 3 T​TS from 111 Elo to 116 for only three times the money. ElevenLabs' price ladder is wider than most comparisons of this matchup admit.

What the 111 points look like in a product

An Elo gap of that size is not subtle. In blind side-by-side listening it is the difference a listener notices on the first sentence, not the fifth — comparable to the gap between Eleven v3 and Eleven v4 within ElevenLabs' own line (1,182 vs 1,328). Cartesia's Sonic 3.6, at 1,289, sits in between and is the model that most directly undercuts the "you get what you pay for" reading; it beats Voxtral by 72 and costs $49, half of Eleven v4.

Where the gap concentrates is delivery control on expressive material. ElevenLabs' v4 generation is built around audio tags — inline instructions for emotion, pacing and emphasis embedded in the text — and around a voice library deep enough that most teams never clone at all. Mistral's Voxtral 3 T​TS offers prompt-level control of emotion, pacing and emphasis plus phoneme-level pronunciation and a vocabulary library. Both expose direction. ElevenLabs has had two generations of public feedback on how those controls behave at the edges, and Voxtral 3 T​TS has had one day.

Where the gap narrows is anything in a European language that is not English. Mistral's model was built multilingual-first, with 50+ dialects claimed and all 24 official EU languages covered across its speech family. ElevenLabs covers 74 languages, which is broader in count, but breadth of count and quality per language are different claims, and neither vendor publishes per-language Elo. If you are shipping Dutch, Polish or Portuguese, neither scoreboard tells you the answer; you have to listen.

What $80 a million characters is actually buying

This is the part of the comparison where the price difference stops looking like a markup. ElevenLabs' per-character rate is bundled into a platform: a shared voice marketplace you can license from rather than clone, professional voice cloning from roughly 30 minutes of audio with a consent workflow, dubbing, sound effects, a conversational-agent stack, and several years of client libraries and integrations. None of that appears on a leaderboard, and all of it is work a team would otherwise do.

Mistral's bundle points in a different direction. Voxtral 3 T​TS ships alongside a LiveKit-based voice-agent SDK with SIP telephony, WebRTC and IVR support, a claimed agent turn under one second, and a per-minute agent price of $0.04 with Mistral Small 4 as the reasoning model. It also ships with an explicit sovereign-deployment pitch: EU, US and global endpoints, self-deployment on private infrastructure, and a stated concern for where audio is processed and whether it is retained. For a European enterprise or a public-sector buyer, that is a procurement answer ElevenLabs does not currently match in the same form.

Put simply: ElevenLabs sells breadth of product, Mistral sells control of deployment. A team that needs dubbing, sound effects and a licensed voice catalogue is not choosing between these two at all. A team that needs one voice, in one language pair, running in a data centre it controls, is paying for a great deal of ElevenLabs it will never switch on.

The migration question, honestly

The tempting framing is "switch from ElevenLabs and save 84%". Run the numbers before believing it, because the saving is only real for workloads that can absorb the quality regression and the missing features.

• Pure narration at scale, single language, no licensed voices — the saving is real. A thousand hours a month is roughly $4,320 on Eleven v4 against $700 on Voxtral 3 T​TS, and the listener is not doing A/B comparisons.

• Anything with a licensed celebrity or brand voice from the ElevenLabs library — the migration is not possible. Voxtral 3 T​TS has no marketplace; you bring 30 minutes of audio and produce your own adaptation.

• Anything user-facing where the customer uploads a short clip and expects to hear themselves — ElevenLabs clones in about 30 seconds and Voxtral, per Mistral's documentation, does not offer an instant clone at all.

• Anything already using dubbing, sound effects or the conversational stack — the effective comparison is not $80 against $13, because leaving means replacing products that were never on the same invoice line.

The realistic pattern is a split: keep ElevenLabs for the parts of the product where a human is listening closely and may compare, and move bulk rendering to the cheaper model. Running that split cleanly is a routing problem rather than a procurement one. OrcaRouter exposes 200+ models behind a single OpenAI-compatible key with provider list price passed through at 0% markup, so a house-voice narration job and a customer-facing clone can point at different backends without a second contract — and if one endpoint degrades mid-render, automatic failover keeps the job moving. As of this writing neither Voxtral 3 T​TS nor ElevenLabs is in OrcaRouter's catalogue; both are reachable through their vendors' own APIs and several third-party platforms, and the routing design is what you build now so the switch is a config change later.

What would change this verdict

Three developments, none of which has happened yet.

The first is a latency number. Mistral has not published time-to-first-audio for Voxtral 3 T​TS. The only latency figures in the launch material belong to the realtime transcription model (sub-500ms) and the agent loop (under one second). Artificial Analysis reports a 112 characters-per-second median throughput, which is faster than Eleven v3's 56 and slower than Sonic 3.6's 145, but throughput is not startup time. If Voxtral 3 T​TS turns out to stream with a first-audio figure near Cartesia's, it becomes a candidate for interactive work and ElevenLabs' price looks worse. If it does not, the streaming category stays closed to it and the $13 is a batch-rendering price only.

The second is independent language-level evaluation. The board scores are aggregate; a model can hold 1,217 Elo overall and be far better or far worse in Italian. Neither vendor publishes the breakdown, and for a European buyer that is the number that matters most.

The third is endurance. Arena clips are short. Mistral's materials do not state a maximum generation length for the T​TS endpoint, and long-form coherence — the thing ElevenLabs' narration heritage is built on — is untested for Voxtral 3 T​TS at any length that would matter to an audiobook producer.

Until those land, the decision is unusually clean. If your product's voice is judged by a customer on a short clip, pay ElevenLabs, and pay for Eleven v4 Turbo at $40 rather than the $80 flagship. If your product's voice is generated in bulk, in a European language, on infrastructure you control, and replayed rather than performed live, Voxtral 3 T​TS at $13 is a defensible bet with a known quality ceiling — and the ceiling is 111 points below the incumbent, which is a number you can plan around rather than a surprise you discover in review.