Mistral and ElevenLabs race to build the AI that answers the phone

On the same day, Mistral released Voxtral 3, a family of speech models plus a kit for voice agents, and ElevenLabs released Eleven Realtime, its first models that take in speech and talk back.

By Roman Aglaganov22 OCT 2026Europe

Illustration: two grey sound waves meeting in the middle, where a single black wave peaks with a red point — Illustration: AI Pulse.

Voice agents, AI that can hold a phone conversation and get something done, are one of the busiest corners of AI. On 20 October two European companies released new systems for building them.

Mistral: Voxtral 3

Mistral’s Voxtral 3 is a set of speech-to-text and text-to-speech models plus a software kit that joins listening, reasoning, using tools and speaking into one real-time pipeline.

Listening: Voxtral Small Transcribe 3 covers more than 50 languages, including all 24 official EU languages, tells speakers apart, and can be primed with key terms such as product names. Mistral says it beats Whisper, ElevenLabs’ Scribe v2 and Gemini 3.5 Transcribe, a claim based on its own tests. A smaller 4-billion-parameter version costs $0.0035 a minute, and a real-time version transcribes in under half a second in more than 35 languages.

Speaking: Voxtral Mini TTS 3 adjusts emotion, pace and emphasis to the conversation, in more than 35 languages and 50 dialects, for $13 per million characters. Mistral says it can adapt to a speaker’s voice and style from 30 minutes of audio.

Mistral offers the models through its EU-hosted service or for companies to run on their own servers, so they control where audio goes and whether it is kept.

ElevenLabs: Eleven Realtime

ElevenLabs, best known for synthetic voices, released two speech-to-speech models: v1 Mini, a cheaper model for natural conversation, and v1 Max, for jobs that need reasoning and tools, such as customer support. Each bundles speech recognition, turn-taking, a language model and voice in one system, so a developer picks one model rather than wiring four together.

In independent tests by Artificial Analysis, Max scored 98.8% on the Big Bench Audio reasoning test and started answering in about 0.57 seconds. ElevenLabs says it supports more than 90 languages and is compatible with OpenAI’s Realtime API, so existing code needs only a new address and model name.

Why it matters

The two take different routes: Mistral sells parts you can run yourself, ElevenLabs a managed system that does it all. Both are betting that call centres and service desks will be among the first places where AI agents do real work at scale, and that speed and natural conversation will decide who wins.