ElevenLabs launches Eleven Realtime: two speech-to-speech models for live voice agents

• Who: ElevenLabs, the AI audio company, via its own product announcement.
• What: Eleven Realtime — two speech-to-speech models, v1 Mini (natural conversation, lower cost) and v1 Max (complex tasks, multi-step tool calls).
• When: Released October 20, 2026 — the same day Mistral launched its Voxtral 3 voice-agent stack.
• Benchmark: v1 Max scores 98.8% on Big Bench Audio with a 0.57-second response time (vendor figures; ElevenLabs says Artificial Analysis ran the evaluations independently).
• Languages: Real-time conversation in more than 90 languages, per the company.
• Access: ElevenAgents platform and an OpenAI-compatible Realtime API; priced per minute of connected audio plus token costs, Max higher per token. Max needs a token limit of at least 1,024 and supports Zero Retention Mode and HIPAA; it is not offered in the India and Singapore data-residency regions.

ElevenLabs has joined the end-to-end voice race. On Tuesday, the AI audio company released Eleven Realtime, two speech-to-speech models — v1 Mini for natural conversation and v1 Max for complex, tool-calling work — designed to let developers ship live voice agents with one model call instead of a stitched-together pipeline. It landed within hours of Mistral launching its Voxtral 3 voice-agent stack: October 20 is officially a voice-agent day.

One model instead of a pipeline

Most voice agents today are assembled from parts: one system transcribes speech, a language model drafts a reply, and a text-to-speech engine reads it out. ElevenLabs says Eleven Realtime folds those pieces into a single system, with speech recognition, turn-taking, the language model, and the voice tuned together. The pitch is straightforward: a voice agent that sounds natural without the developer hand-tuning each stage.

The models are built for messy, real conversations — interruptions, quick back-and-forth, the filler sounds people make. According to The Decoder, the smaller v1 Mini can tell the difference between a speaker who has finished talking and one who is just saying "mm-hmm" — one of the small, hard problems in making voice agents feel human.

Mini for conversation, Max for the phones

The two models target different jobs. Mini is the cheaper, conversation-first option: AI assistants, tutors, interactive characters. Max adds reasoning for harder work — customer support, lead qualification, multi-step workflows with chained tool calls — and is the only one of the two to support Zero Retention Mode and HIPAA, which puts Mini off the table for regulated health workloads.

ElevenLabs recommends starting with Mini and moving to Max only when an agent follows long instructions, relies on chained tool calls, or must meet stricter compliance rules.

The numbers — and the asterisks

ElevenLabs claims v1 Max scores 98.8 percent on the Big Bench Audio benchmark with a 0.57-second response time, and says no model on the Artificial Analysis speech-to-speech leaderboard is currently both faster and more accurate. These are the company's figures — though ElevenLabs says Artificial Analysis ran the evaluations independently. Either way, the battleground has shifted: turn latency and interruptibility now matter more than raw transcription accuracy.

An OpenAI-compatible API as a strategy

The second way in is tactical. Alongside the ElevenAgents platform, the models are available through a Realtime API that mirrors the format of OpenAI's own: ElevenLabs says developers can point an existing OpenAI Realtime client at its endpoint by changing the endpoint and model name, with client and server events arriving in the same shape so existing handlers and logging keep working. It is a direct bid for developers who already built on that standard — and puts ElevenLabs in open competition with OpenAI on all-in-one realtime voice.

Pricing is per minute of connected audio plus token costs, with v1 Max costing more per token than v1 Mini; current rates are on ElevenLabs' pricing page. The company also pitches conversation in more than 90 languages, and agents built on ElevenAgents can deploy to phone, a web widget, WhatsApp, and contact-center platforms.

The fine print worth reading

There are limits. The Realtime API connects over WebSocket only at launch, so apps using WebRTC have to switch first. Agents that use a custom LLM, Amazon Bedrock, or workflow nodes that override managed settings must stay in the older Custom mode. v1 Max requires a token limit of at least 1,024 and is not offered in the India and Singapore data-residency regions. And the rollout is gradual — some workspaces will not see the Realtime option yet.

Still, the timing is the story. Two voice labs shipped full agent stacks within hours of each other, each taking the opposite architectural bet: Mistral's Voxtral 3 is a composable stack of transcribe, speech, and agent-SDK pieces with a European sovereignty pitch, while ElevenLabs bets that one tuned end-to-end model is the simpler thing to ship. Developers now get to vote with their call scripts.