The most interesting line in ElevenLabs’ new launch isn’t the benchmark score. It’s the migration path. Eleven Realtime follows OpenAI’s Realtime API spec, so a developer can switch by changing the endpoint and the model name. That is a direct shot at a competitor’s customer base, and it’s a smart one.
ElevenLabs launched Eleven Realtime as a family of speech-to-speech models for building voice agents. There are two variants. Eleven Realtime v1 Mini is aimed at AI assistants, teachers and interactive characters. Eleven Realtime v1 Max is aimed at customer support, lead qualification and multi-step workflows. Both support more than 90 languages and can use any of the 17,000-plus voices in the ElevenLabs library.
The pitch is simplicity. The models fold speech recognition, turn-taking, voice activity detection and speech generation into one system. Anyone who has stitched those pieces together knows where it hurts: deciding when to respond after a pause, handling interruptions, and keeping voice output in sync with the language model behind it. ElevenLabs says developers shouldn’t have to become configuration experts to get a natural-sounding agent.
The numbers, from independent tests by Artificial Analysis, show a real split between the two tiers. On Big Bench Audio, which tests reasoning through spoken questions, Max got 98.8% right against 75.9% for Mini. On tool use, Max completed 87.1% of tasks and Mini 60.3%. In the Speech Agent Arena, which uses blind live conversations judged on human preference and task success, Max scored 998 and Mini 988. Max also had a time to first audio of 0.57 seconds versus 0.62 for Mini.
But Mini wins in one place. On Full-Duplex-Bench, which measures pauses, interruptions, turn-taking and backchannels, Mini scored 95.8% and Max 94.1%. So the cheaper model is slightly better at sounding like a person in conversation, while the bigger one is better at thinking and acting. That’s a sensible division. A tutor or game character needs rhythm. A support agent that books refunds needs tool calls that work.
The compatibility angle matters most for the market. OpenAI’s Realtime API has become the reference interface for voice agents, and many teams have built on it. Matching the spec lowers switching costs to nearly zero, which forces a comparison on quality, voice choice and price. There is one catch: the API connects over WebSockets and works with OpenAI’s SDK, so anyone using WebRTC will need to move to WebSockets first.
Developers who want more than a model can use ElevenAgents, the company’s agent platform. It covers deployment across phone, web, WhatsApp and contact centre channels, and includes workflows, knowledge bases, guardrails, testing and monitoring. The announcement didn’t give pricing in the details reported, so cost per minute is the first thing to check before committing.
Does it matter? Yes, with caveats. ElevenLabs built its name on voice quality, and its library of 17,000 voices is a real differentiator against OpenAI, Google and smaller players. Moving from text-to-speech into full speech-to-speech agents puts it in direct competition with the labs it once complemented. Benchmark wins are vendor-selected and narrow, though. A 98.8% reasoning score still leaves roughly one in 100 spoken questions wrong, which is a real consideration in customer support. Run your own call transcripts through both tiers before you switch.
