A.I. news · October 21, 2026
ElevenLabs v1 Max: 98.8% on Spoken Reasoning, 0.57s Reply
ElevenLabs launched Eleven Realtime v1 Mini and v1 Max, two Speech to Speech models that each work as a complete voice agent, and says v1 Max got 98.8% on a spoken reasoning test.
What ElevenLabs launched 🎙️
The old way meant picking each piece yourself. Those pieces include speech recognition, voice activity detection, turn-taking, the underlying LLM and text to speech. ElevenLabs calls building that stack a project in itself.[1]
Both models support more than 90 languages. They can use any of the 17,000+ voices in the ElevenLabs library.[1]
Our take: think of the old stack as ears, a sense of when to jump in, a brain and a voice. This launch bundles all four into one model.[1]
Mini or Max? 🔎
Eleven Realtime v1 Mini is for natural conversations at low cost. ElevenLabs says it suits A.I. assistants, teachers and interactive characters. It also says Mini will be sufficient for most use cases.[1]
Eleven Realtime v1 Max is built for complex tasks. ElevenLabs names customer support, lead qualification and multi-step workflows.[1]
ElevenLabs recommends Max for heavy tool calling. On tool-use tasks, v1 Max completed 87.1% and v1 Mini completed 60.3%.[1]
The benchmark numbers 📊
ElevenLabs says Artificial Analysis evaluated both models independently across three public benchmarks.[1]
On Big Bench Audio, v1 Max answered 98.8% correctly and v1 Mini answered 75.9%. V1 Max starts speaking in 0.57 seconds on average. V1 Mini takes 0.62 seconds.[1]
ElevenLabs says accuracy and speed are usually a trade-off. It claims v1 Max excels at both.[1]
On Full-Duplex-Bench, v1 Mini scores 95.8% and v1 Max scores 94.1%. That test covers pauses, turn-taking, interruptions and backchannels.[1]
In the Speech Agent Arena, v1 Max scored 998 and v1 Mini scored 988.[1]
Our take: Mini scoring a bit higher on Full-Duplex-Bench could mean a chatty assistant doesn't need the bigger model, if the result holds.[1]
How to try it 🛠️
ElevenLabs says its realtime API is compatible with OpenAI's Realtime API. Trying it means changing the endpoint and model name in your existing code.[1]
It connects over WebSocket and follows OpenAI Realtime specs, so OpenAI's SDK works as is. It is WebSocket only, and it runs separately from ElevenAgents. A migration guide covers the steps and the differences from OpenAI's API.[1]
In ElevenAgents, you create or open an agent and pick a Realtime model from a dropdown. The platform includes channels, workflows, a knowledge base, guardrails, testing and monitoring.[1]
ElevenLabs says small judgment calls, like pause timing and interruptions, are what make an agent feel right. In ElevenAgents, those are adjustable settings. If you need channels or a knowledge base, it says to build on ElevenAgents from the start.[1]
What we don't know yet ⚠️
The New Stack ran a piece headlined: ElevenLabs wants to run your voice agent. Here's what developers give up.[2]
Our take: that headline hints at a trade-off of less work, maybe less control. The big question is whether these results survive real phone lines, accents and frustrated customers.[2]
💡 Our take: what this could mean for you
- If you already use OpenAI's Realtime API, a test may take little more than swapping the endpoint and model name. Check that you can use WebSocket first.
- Match the model to the job: Mini for casual conversation, Max for tool-heavy work like support and lead qualification.
- Test with pretend calls that include interruptions and changes of mind. Confirm pricing and limits before you commit.
Have a business or a business idea? 🚀
Trinity builds a free, personalized website prototype for it — no credit card.
Get my free prototype