Eleven Realtime: Voice Agents for Support and Lead Calls

ElevenLabs launched Eleven Realtime on October 20, 2026. It is a pair of speech-to-speech models for voice agents. Each one handles listening, deciding when the caller has finished, thinking and speaking as a single choice, so you no longer pick a speech recognizer, a language model and a voice separately and then tune how they work together. Eleven Realtime v1 Mini is the default and the cheaper option. Eleven Realtime v1 Max reasons before it answers and is the one ElevenLabs recommends for customer support, lead qualification and multi-step workflows.

This post covers what changed, how to choose between Mini and Max, what a minute costs, and how a small business would set up a phone agent for support or lead calls, including when it is not worth doing.

We read these ElevenLabs pages on October 21, 2026: the launch post Introducing Eleven Realtime, the docs page Realtime models and the API pricing page. Every ElevenLabs fact below comes from those pages. We have not run a production phone line on these models yet.

Until now, a voice agent on ElevenAgents (ElevenLabs' agent platform) was built from parts: a speech recognizer, a language model, a text-to-speech voice and settings for pauses and interruptions. ElevenLabs calls that Custom mode, and it still exists. Realtime mode is the new option. You pick Mini or Max, and ElevenLabs manages the language model, the speech recognizer, the voice model and a few reasoning settings for you.

What you keep is what makes the agent yours: the system prompt and first message, the voice (any of the 17,000+ in the ElevenLabs library), languages (more than 90), tools, the knowledge base, and conversation settings such as how eager the agent is to take its turn and how it handles interruptions. The docs say these work the same way in both modes.

Agents on ElevenAgents can be deployed to phone, web, WhatsApp and contact center platforms. For developers already on OpenAI's Realtime API, ElevenLabs offers a compatible Realtime API, so trying it means changing the endpoint and model name. That API connects over WebSocket only, and it runs apart from ElevenAgents, so phone channels, the knowledge base and the other platform features are not part of it.

ElevenLabs reports results from Artificial Analysis, which tested both models on three public benchmarks. These are vendor-reported numbers, and we have not reproduced them:

• Spoken reasoning questions answered correctly (Big Bench Audio): Max 98.8%, Mini 75.9%
• Average time before the agent starts speaking: Max 0.57 seconds, Mini 0.62 seconds
• Turn-taking, pauses and interruptions (Full-Duplex-Bench): Mini 95.8%, Max 94.1%
• Tool-use tasks completed: Max 87.1%, Mini 60.3%

The tool-use gap is the number that matters for most business calls. A support or lead call is rarely just talk. The agent has to look up an order, check a calendar, find a customer in the CRM or book a slot. If about 4 in 10 tool tasks go unfinished, as Mini's reported score suggests, the caller hears an agent that sounds natural but cannot do the job. That is why ElevenLabs recommends Max for heavy tool calling, and we would start any call that touches your systems on Max.

Mini fits calls that are mostly conversation: answering common questions from a knowledge base, giving hours and directions, or taking a message.

Two more rules from the docs decide the choice for some businesses. Mini is not available under Zero Retention Mode or HIPAA, while Max is available under both in the US and EU. Max is not available in the India and Singapore data residency environments. Realtime models are also rolling out gradually, so if the Voice mode setting does not appear on your agent, your workspace does not have them yet.

The docs list cases where you should keep the agent in Custom mode:

1. 1You use your own language model, Amazon Bedrock or Speech Engine.
2. 2You need one specific language model or voice model.
3. 3Workflow nodes in the agent override the model or another managed setting.
4. 4The agent transfers calls to or from agents that are in Custom mode. Realtime agents can only transfer to other Realtime agents, though Mini and Max can transfer to each other.
5. 5The agent is text only.

Point 4 catches teams that already run several agents. If your front desk agent hands billing questions to a separate billing agent, switch both or neither.

The pricing page lists $0.08 per minute of connected audio for both Mini and Max, plus token charges. Mini tokens cost $0.20 per million input tokens and $6.00 per million output tokens. Max costs $1.80 per million input tokens and $9.00 per million output tokens. Each plan includes a number of minutes, and ElevenLabs says those are estimates that vary with conversation length and tool use. Calls above your plan's concurrency limit, which ranges from 4 to 40 concurrent calls across the listed plans, are billed at a burst rate of $0.16 per minute.

A worked example, using only the connected audio rate: a business that takes 75 calls a month averaging 4 minutes uses 300 minutes. At $0.08 a minute, that is $24 before tokens. Token cost depends on how long your prompt is, how much the caller says and how many tools the agent calls, so measure it on test calls before you commit. Your phone line or contact center provider may also charge separately. These numbers come from the published rates, and we have no client results on this model yet.

The bigger cost is usually setup and upkeep: writing the prompt, connecting the tools, testing hard calls and reviewing transcripts each week.

These are designs we would build. None of them describes a client result.

1. 1A lead qualification line. Calls to your sales number reach a Max agent that asks the four or five questions your team always asks, looks up whether the caller is already in your CRM, books a call on the right person's calendar and writes a summary to the CRM record. Anyone who asks for a person, or who fits your best-customer profile, gets transferred during business hours or promised a callback. Our post on real estate lead response time shows why the first minutes after a lead calls matter.
2. 2A support first line. The agent answers order status, return policy and account questions from your knowledge base and a read-only order lookup. Anything that changes money or an account, such as a refund or a cancellation, becomes a ticket for a person to approve. Our post on AI agent permissions and human approval covers where to draw that line.
3. 3An after-hours line. The agent answers routine questions, takes messages and books appointments, and sends anything urgent to the on-call path you already have. For medical practices, our post on after-hours patient calls covers what an agent must not handle.

For each one we would set up the same pieces: a short list of tools, read-only where possible; a clear handoff to a person, with the call summary passed along; a written list of topics the agent must refuse; call recording and transcript review; and a weekly check of cost per call and of the calls the agent got wrong. Before going live we would run test calls with interruptions, background noise, accents and callers who change their mind halfway, using the approach in our pre-production testing checklist.

If you get a handful of calls a week, a good voicemail with a text-back and a person returning calls within an hour may serve callers better than an agent you have to maintain. If most calls need judgment, such as quoting custom work or handling an upset long-time customer, an agent will hand off so often that it adds a step rather than removing one. If your business data lives in spreadsheets and email rather than a CRM or booking system, the agent will have nothing to look things up in, so fix that first. Our post on AI agents vs. automations helps sort which jobs need an agent at all.

Also check the rules for your industry and region on recording calls and telling callers they are speaking with an AI before you launch.

OrchestriAI is not an ElevenLabs partner or reseller. We build AI agent systems and connect them to the CRM, calendar and help desk tools a business already uses through systems integration, with clear limits on what each agent can change. If you want to know whether a phone agent would pay off for your call volume, start with the free automation audit.

OrchestriAI is an independent implementation provider. Product names are used for identification only. OrchestriAI is not affiliated with, endorsed by, or sponsored by ElevenLabs. This article summarizes ElevenLabs' public launch post, documentation and pricing page as of October 21, 2026. Benchmark results are as reported by ElevenLabs. Prices, plan limits and model availability can change, so confirm them on ElevenLabs' site and in your workspace before you plan a budget.