Eleven Realtime: what it is, pricing and how to try it
2026-10-22 · 4 min read
Eleven Realtime is a new pair of speech-to-speech models from ElevenLabs, built for voice agents that answer phones, tutor students or qualify leads. There are two versions, v1 Mini and v1 Max, and both are live now inside ElevenAgents and through an API that copies OpenAI's Realtime API, at $0.08 per connected minute plus token costs.
The launch posts
@ElevenLabs announced the family on October 20 with a 71-second demo video. Their argument: voice agents usually make you pick between a fast reply and a correct one, and "Eleven Realtime gives you both." The company says the models start talking in about half a second and still get spoken reasoning questions right.
A follow-up post from @ElevenLabs explains what Mini is for. It knows when you have finished speaking, recovers when you interrupt, and keeps going when you say "mm-hmm." Those three habits decide whether a phone agent feels like a person or like a voicemail tree with opinions, so I'm glad they called them out.
Mini vs Max: what each one does
ElevenLabs' Realtime page and pricing page give these numbers. The benchmarks come from ElevenLabs itself, so read them as claims until someone outside the company tests them.
| v1 Mini | v1 Max | |
|---|---|---|
| Built for | Assistants, tutors, characters | Support, lead qualification, multi-step tasks |
| Time to first audio | 620 ms | 570 ms |
| Reasoning and tool calls | Lowest latency and cost, simpler tasks | Reasons before it answers, multi-step tool use |
| Vendor benchmark | Full-Duplex-Bench 95.8% | Big Bench Audio 98.8%, task success 87.1% |
| Connected audio | $0.08 per minute | $0.08 per minute |
| Input tokens (per 1M) | $0.20 | $1.80 |
| Output tokens (per 1M) | $6.00 | $9.00 |
Both versions bill $0.16 a minute for burst traffic above your concurrency limit. ElevenLabs says the models handle more than 90 languages and can use any of its 17,500+ voices.
Max is a little faster to start talking, which is odd since it also does more thinking. The bigger gap is tool use: Max can look something up or book something in the middle of a call. Mini is the cheap, chatty one.
How to try it
The quickest route is the ElevenAgents dashboard. Per the Realtime models docs:
- Open or create an agent and go to its Agent tab.
- Set Voice mode to Realtime. That picks Mini by default.
- To use Max, choose it from the Model list, then publish.
Once you're in Realtime mode, ElevenLabs picks the LLM, the speech recognition and the voice model for you. You still write the system prompt and the first message, and you still attach tools and a knowledge base. The model selector shows an estimated cost per minute before you publish.
Developers who already built against OpenAI's Realtime API can switch by changing the endpoint to wss://api.elevenlabs.io/v1/convai/realtime and the model name to eleven_realtime_v1_mini or eleven_realtime_v1_max. The catch is that the API path leaves out the ElevenAgents extras like phone numbers, knowledge bases, guardrails and testing, and it doesn't support WebRTC at launch.
A test prompt that tells you something useful: ask it to book a Tuesday afternoon appointment, then cut it off halfway through its reply and change the day. A good agent drops what it was saying and follows you, while a weaker one plows through to the end of its sentence before it notices.
The fine print
A few limits from the docs matter for small businesses:
- Rollout is gradual. If the Realtime option is missing, your workspace doesn't have it yet.
- Mini is not available under HIPAA or Zero Retention Mode. Max is, in the US and EU. A dental or medical office that needs HIPAA has to use Max.
- You can't pair Realtime with a custom LLM or Amazon Bedrock, and it doesn't do text-only agents.
My read
ElevenLabs has had the voices for a while. What it lacked was the part that makes a phone agent feel quick, and this is its answer to OpenAI's realtime models and to Mistral's cheaper stack, which we covered in Voxtral 3 pricing. It landed the same week as the company's video tool, ElevenLabs Chimera, and both launches lean on the same voice models, so that's plainly where ElevenLabs is putting its chips.
For a local service business, $0.08 a minute plus tokens means an AI agent picking up overflow and after-hours calls costs pennies per call. Whether it should is a separate question. Before you hand your phone line to any model, write down what callers actually ask and which answers need a human. Our comparison of answering services vs missed-call text-back is a good place to start. If you'd like a second pair of eyes on where a voice agent fits in your intake, New Face Design's free process audit maps that out before you spend anything.