ElevenLabs Realtime Explained: Voice Agents in One Model

What used to take five parts now takes one choice

Before Realtime, building a voice agent meant picking and tuning five pieces yourself: speech recognition, voice-activity detection, turn-taking, a language model, and text-to-speech, then fixing the awkward pauses where they handed off to each other. ElevenLabs calls that setup Custom mode, and it still exists.

Realtime mode replaces it with a single model selection. ElevenLabs manages the language model, the speech recognizer, the voice model, and reasoning settings for you. What stays yours is what makes the agent yours: the system prompt and first message, the voice, the languages, the tools, the knowledge base, and settings like how eagerly the agent takes its turn and how it handles interruptions.

The practical payoff is speed. According to ElevenLabs, v1 Max starts speaking 0.57 seconds after a question on average, and v1 Mini follows at 0.62 seconds — fast enough that a conversation feels like a conversation, not a series of voice messages.

Mini vs Max: which one fits what

The benchmark scores come from third-party evaluator Artificial Analysis, as reported by ElevenLabs in October 2026 — vendor-reported numbers, not independent long-term testing, so treat them as a launch snapshot. Still, the gap tells you the picking rule ElevenLabs itself recommends: start with Mini, and move to Max only when your agent follows long instructions, chains tool calls, or must meet stricter compliance rules. Only Max supports HIPAA and Zero Retention Mode, so regulated use cases rule Mini out immediately.

A representative of ElevenLabs put the ambition plainly in launch coverage: “Through Eleven Realtime, which supports more than 90 languages, including Korean, we plan to help Korean businesses and developers build voice agents that can converse naturally with customers.” That sentence is the whole strategy in one line — voice agents as a normal business tool, not a developer demo.

How you actually use it

There are two routes in, and they suit different people:

1. ElevenAgents (the no-code route). Open an agent, switch its voice mode to Realtime, and pick Mini or Max. Your prompt, voice, tools, and knowledge base keep working as before. Agents can deploy to phone lines, a web widget, WhatsApp, and contact-centre platforms such as Zendesk and Genesys.
2. The Realtime API (the developer route). If you already have code built on OpenAI's Realtime API, trying ElevenLabs means changing the endpoint and model name — OpenAI's own SDK works without modification, because the API follows OpenAI's Realtime specification over WebSocket.

Two honest caveats before you plan around it. First, the API is WebSocket only at launch: there is no WebRTC version yet, so a browser-based app built on WebRTC has to change its transport layer first. Second, launch coverage notes that no price was listed in the launch post — confirm pricing in your ElevenLabs workspace before budgeting a phone line on it.

The trade-off nobody should skip

Realtime's convenience has a price that is not measured in dollars: control. When one model runs the whole loop, you can no longer swap out a single piece — if a better or cheaper language model ships next month, you wait for the vendor. Critics have also flagged that the bundled model removes the text transcript that most enterprise safety and compliance tooling is built around. For a customer-support line, that is not a footnote; it decides whether your existing monitoring still works.

That does not make Realtime a bad choice. It makes it a choice with a clear shape: maximum simplicity and speed, minimum ability to tinker. If you want the opposite shape — full control, no subscription, your audio staying on your own machine — the open-source route we covered in our Audio8 voice cloning guide is the mirror image. And if you want to understand the underlying technology first, our speech-to-speech explainer walks through how speech-in, speech-out models work in plain English.

What this means for you right now

If you are a business owner, not a developer, the honest read is this: you do not need Realtime today, but you should know it exists, because the support line or booking assistant you buy next year will very likely run on something shaped exactly like it. If you are a creator who just wants a voiceover today, none of this changes your workflow — you still need text turned into speech, not a live agent.

And that part is already free and instant. Paste your script into TextToSpeechH, pick a voice, and download the MP3 — no sign-up, no card, no API key. Voice agents are the future of phone calls; a free voiceover in your browser is the present of everything else.

One step further: if you want the finished-product version of this idea — a no-code agent that answers and qualifies sales calls inside the platform your team already calls from — our Convoso Voice AI Agent guide covers that route, including the early results Convoso published at launch.

If you hit ElevenLabs' free limits while experimenting, our ElevenLabs alternatives guide compares seven free tools with bigger limits and no sign-up.

And if you are arriving here because your voice-agent platform is disappearing, you are not alone: Hume's EVI voice API shuts down on November 13, 2026. Our Hume AI alternatives guide covers what replaces EVI — Realtime included — and what replaces its Octave text-to-speech, with the free options first.