ElevenLabs has launched Eleven Realtime, two models that take in a caller's voice and answer in speech, with the listening, the timing of turns, the thinking and the voice all built as one piece. Mini costs less and is meant for assistants, tutors and characters. Max costs more and is meant for support calls, lead qualification and anything with several steps.
The timing used to be the hard part
Until now, building a phone agent that did not feel awkward meant choosing a speech recognizer, a system that detects pauses, a language model and a voice, and then tuning them together. The tuning decided small things: how long to wait after a pause, and what to do when the caller talks over the agent. People leave about a fifth of a second between turns across languages, so a delay of a second reads as a stranger who is not listening.
Both new models start speaking in about half a second on average. In the ElevenAgents builder, choosing one is a single dropdown, and agents that already exist do not change unless their owner switches them.
Sounding right and finishing the job are different skills
The two models are almost level on conversation flow, and Mini is slightly ahead. The gap shows up when the agent has to act. On tasks that need tools, such as checking a calendar and then moving a booking, Mini completed 60.3% and Max 87.1%. On spoken reasoning questions it was 75.9% against 98.8%.
Take a physiotherapy clinic that gets 40 calls a day from patients wanting to move or book a visit. If those calls behave like the tasks in that test, Mini leaves about 16 of them unfinished and Max about 5. Each unfinished call is a patient who tries again, or a receptionist who has to call back. Over a 22-day month, that is roughly 350 unfinished calls against 110.
The sensible setup splits the work: Mini for questions about opening hours and directions, Max for any call that changes a calendar or an order. The other half of the setup is a handover. Every agent will miss some calls, so the clinic decides what the agent says to pass the patient to a person, and whether that happens before the caller gets frustrated.
ElevenLabs already sold an AI receptionist for small businesses in September. The news here is that the model underneath is no longer a project to assemble, so the sensible question for a small business has moved from whether the voice is good enough to whether the work gets done.
A second company now fits the same plug
OpenAI already sells speech-to-speech models through its Realtime API, and its upgraded one came out in August 2025, so the idea is not new. What is new is that ElevenLabs made its interface follow OpenAI's. A team already using OpenAI's version changes the web address and the model name, and the rest of its code keeps working.
When switching costs two lines, vendors have to compete on price and finished jobs every month, because their customers can leave in an afternoon. I expect the price of a voice-agent minute to fall over the next year, and the stronger tier to be where the real difference between vendors shows.
There are limits. The raw API connects only over WebSocket, not WebRTC, and it does not include phone numbers, WhatsApp, a knowledge base or guardrails; those come with the ElevenAgents platform.
Whichever vendor a business picks, the voice will sound fine. The useful question is what share of calls end with the job done, and what happens to the ones that do not.