ElevenLabs' Realtime Max Hits 99% on Speech Reasoning and Ranks Ninth With Callers. Mistral's "State-of-the-Art" Transcribe Is Third Behind Microsoft's Cheaper Model.
Two voice launches landed on October 20. ElevenLabs shipped Realtime v1 Mini and Max, speech-to-speech agents compatible with OpenAI's Realtime API; Mistral shipped Voxtral 3 transcription, TTS, and a voice agent preview and called its transcribe model state of the art. Artificial Analysis disagrees with the loudest claims: Max scores 98.8% on Big Bench Audio and 997 Elo on the Speech Agent Arena, under the 1,000 anchor; Voxtral Small Transcribe 3 is third on word error rate behind Fun-ASR and Microsoft's MAI-Transcribe-2, which is also cheaper and faster. We read both vendors' pricing and docs and AA's boards.
By T.W. Ghost
Six days after Grok Voice 3.0 took first place on Artificial Analysis's speech board and five days after OpenAI moved ChatGPT to GPT-Live-1.1, two more voice products launched on the same Tuesday, October 20, 2026. ElevenLabs shipped Realtime v1 in two sizes, Mini and Max, a speech-to-speech agent stack compatible with OpenAI's Realtime API, rolling out workspace by workspace. Mistral shipped Voxtral 3: transcription in three tiers, text-to-speech, and a voice agent in public preview, with its Small Transcribe model billed as "state-of-the-art."
Both launches came with charts. Both charts leave out the models that beat them. ElevenLabs' Max scores near the top of AA's speech reasoning test and ninth with callers on the arena board. Mistral's transcribe is third on AA's word-error-rate board behind two models its chart does not show, one of which is cheaper and much faster. We read ElevenLabs' blog, pricing, and model docs, Mistral's post and pricing, and AA's speech-to-speech, arena, and speech-to-text boards. Here is what held up.
ElevenLabs Realtime v1: Two Sizes, Two Compliance Stories
| Realtime v1 Mini (default) | Realtime v1 Max | |
|---|---|---|
| Launch | October 20, 2026 | October 20, 2026 |
| API price | $0.20 in / $6 out per 1M tokens | $1.80 in / $9 out per 1M tokens |
| Plus | $0.08 per minute of connected audio; $0.16 per minute burst above concurrency | Same |
| Vendor claims | 96% conversational dynamics; 620 ms to first audio | 99% speech reasoning; 570 ms to first audio |
| AA Big Bench Audio | lower | 98.8% (third) |
| AA time to first audio | not recorded here | 0.57 s |
| AA Speech Agent Arena (Elo) | 987 | 997 (about ninth) |
| Compliance | Not under zero-retention or HIPAA | ZRM and HIPAA in US and EU; not in India or Singapore residency |
| Languages | 90+ | 90+ |
Price it as minutes plus tokens. The per-token rates are the headline; the $0.08 per connected minute, and $0.16 above your concurrency, is the line that scales with call volume. A quiet call still runs the meter.
It is a managed stack, not one model. ElevenLabs' docs describe Realtime as a cascaded system of ElevenLabs-managed components, not a single fused model where you pick the LLM and the voice. That matters for anyone comparing it to a true speech-to-speech model.
The arena is where the 99% goes to be humbled. AA's Big Bench Audio puts Max at 98.8%, third, which ElevenLabs rounds to 99%. On AA's Speech Agent Arena, where humans rate calls, Max sits at 997 Elo, below the 1,000 anchor set by GPT-Realtime-1.5 and far below Grok Voice 3.0 at 1,064 and Gemini 3.8 Live at 1,077 as of October 23. ElevenLabs' blog printed 998; AA shows 997. A model can reason about audio very well and still not be the one callers prefer.
Mistral Voxtral 3: Cheap, Fast Enough, and Not First
| Voxtral 3 (public preview, Oct 20) | |
|---|---|
| Mini Transcribe 3 | $0.0035 per minute |
| Small Transcribe 3 | $0.004 per minute |
| Mini Transcribe Realtime 3 | $0.008 per minute |
| TTS 3 | $13 per 1M characters |
| Voice agent (with Mistral Small 4) | $0.04 per minute |
| Self-deployment | "By arrangement"; no Voxtral 3 weights on the public Hugging Face listing we checked |
| Naming | The post says "Voxtral Speech 3" and "Mini TTS 3" in places; the docs and pricing say "Voxtral TTS 3" |
AA's word-error-rate board, non-streaming (AA-WER v2): Fun-Realtime-ASR-preview 1.73%; Microsoft MAI-Transcribe-2 2.04%; Voxtral Small Transcribe 3 2.09%. Mistral's chart compares against Gemini 3.5 Transcribe, Scribe v2, and Whisper, and omits the two models above it.
AA's price and speed columns: MAI-Transcribe-2 at $1.67 per 1,000 minutes and about 304 times realtime; Voxtral Small at $4.00 per 1,000 minutes and about 65 times realtime. On AA's measures Microsoft's model is more accurate, cheaper, and faster. "State-of-the-art" is Mistral's word for third.
Three Honest Observations
- Reasoning board and caller preference are different tests. Max's 98.8% is real and so is its 997. Buy for the one your product needs.
- The compliance split is inside one product. Mini cannot do zero retention or HIPAA; Max can, in the US and EU only. Read the model docs before the pricing page.
- Neither vendor's chart shows the leader. That is the pattern of the week, and the reason the house rule is to read AA before the launch post.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Realtime v1 Mini and Max launched Oct 20; OpenAI Realtime-compatible; 90+ languages; prices as listed | Verified (ElevenLabs blog, pricing, docs) |
| Max "99% speech reasoning," 570 ms | Vendor claims; AA measures 98.8% and 0.57 s |
| Max Speech Agent Arena Elo 997, below the 1,000 anchor | Verified (AA, Oct 23 read); ElevenLabs printed 998 |
| Mini not under ZRM or HIPAA; Max US/EU only for those | Verified (ElevenLabs model docs) |
| Voxtral 3 prices as listed; public preview Oct 20 | Verified (Mistral post and pricing) |
| Voxtral Small Transcribe 3 is state of the art | Contradicted on AA's board (third, behind Fun-ASR and MAI-Transcribe-2) |
| MAI-Transcribe-2 cheaper and faster per AA | Verified (AA STT board) |
| Voxtral 3 open weights | Unverified; none on the public Hugging Face listing we checked |
| Mini task-success 60.3% | Unverified (not re-confirmed on AA) |
Who Should Use It
Use ElevenLabs Realtime Max if your agent needs to reason about audio and you need HIPAA or zero retention in the US or EU, and price it in minutes. Use Mini for cost-sensitive, non-regulated traffic. Test both on the arena-style question, whether callers prefer it, because that is where the board says the gap is.
Use Voxtral Small Transcribe 3 if you want a cheap European transcription option and have checked the board. For a transcription bake-off, include MAI-Transcribe-2 and Fun-ASR, not only Whisper and Scribe; Mistral's chart will not do it for you.
The frontier is elsewhere: on AA's speech-to-speech board the top is Grok Voice 3.0 and the Gemini and OpenAI live models; on transcription it is Fun-ASR and Microsoft. Neither of today's launches changed the leaders, and both launches claimed to. Our model comparison has the current standings side by side.
Sources
- ElevenLabs: Realtime
- ElevenLabs: API pricing
- ElevenLabs docs: Realtime models
- Mistral: Voxtral 3
- Mistral docs: inference pricing
- Artificial Analysis: speech to speech
- Artificial Analysis: Speech Agent Arena
- Artificial Analysis: speech to text, non-streaming
*Not sure whether your voice agent needs reasoning scores or caller preference? Take the free 2-minute quiz and get matched. Then read about the model that actually sits first on the speech board: Grok Voice 3.0.*