ElevenLabs Eleven Realtime: Max vs Mini (2026)

ElevenLabs has launched Eleven Realtime, its first end-to-end speech-to-speech models built for voice agents. There are two of them: v1 Max, the flagship aimed at high-accuracy work like customer support, and v1 Mini, a smaller, cheaper model for natural everyday conversations. Both arrived on October 20, 2026 — the very same week as Mistral's Voxtral 3 voice-agent stack, which makes this the busiest week in voice AI in a long time.

If you are building — or buying — a voice agent, this post answers the practical question: what Max and Mini actually are, how they differ, what they cost, and which one you should try first. Specs below are as of October 21, 2026.

What Eleven Realtime is

Traditional voice agents are assembled from parts: a speech-recognition model that transcribes what the caller says, a language model that decides what to say back, and a text-to-speech voice that speaks it. Each part adds latency and its own failure modes.

Eleven Realtime collapses that pipeline into one model choice. You pick v1 Max or v1 Mini and get a configured voice agent: listening, reasoning, and speaking bundled into a single API surface. The company positions it as "voice agents in one model call." Both models support more than 90 languages and access to ElevenLabs' library of over 17,000 voices, according to launch coverage.

There are two ways to build on them: through the ElevenAgents platform (ElevenLabs' no-code/low-code agent builder) or through an OpenAI-compatible Realtime API, which matters a lot if you already built on OpenAI's standard — you can swap the endpoint and model name and trial ElevenLabs side by side instead of rewriting your integration.

Max vs Mini: how they differ

| | v1 Max | v1 Mini | |---|---|---| | Best for | Customer support, complex multi-step tasks, tool-calling | Natural everyday conversations | | Accuracy | 98.8% on Big Bench Audio (company claim) | Lower, not the flagship for accuracy | | Response speed | ~0.57 seconds (company claim) | Fast enough for casual chat | | Cost | Higher per minute + tokens | Lower per minute + tokens | | Standout feature | Top-left of the Artificial Analysis speed-vs-accuracy chart (company claim) | Detects turn-taking: knows when you finished vs. just saying "mm-hmm" | | Compliance | HIPAA + Zero Retention Mode supported | No HIPAA / Zero Retention support | | Access | ElevenAgents + OpenAI-compatible Realtime API | Same |

A few things to note about the table above:

Every benchmark number is ElevenLabs' own claim. The 98.8% Big Bench Audio score and the 0.57-second response time come from the company's launch materials. No independent benchmarks have been published yet, and the Artificial Analysis leaderboard comparison is the company's own reading of the chart. Treat the numbers as a vendor claim, not a fact.

Mini's turn-taking is the interesting design decision. Interruption handling — knowing when someone is done talking versus back-channeling with "mm-hmm" — is one of the hardest parts of voice agents to get right, and it's where cheaper models usually feel clunky. Giving the cheaper model the smarter conversational instinct is a sensible segmentation: Max for when accuracy decides outcomes, Mini for when the conversation just needs to feel human.

Only Max supports HIPAA and Zero Retention Mode. If you handle health data or regulated records, Mini is off the table entirely. That is a pricing and compliance gate worth knowing before you prototype on Mini and discover the compliance gap later.

What it costs

Pricing is usage-based: per minute of audio plus token costs, with v1 Max costing more than v1 Mini. ElevenLabs displays an estimated per-minute cost in the model selector so you can sanity-check before a build.

For context on ElevenLabs' general pricing (as listed in late September 2026): the Free tier is $0 with 10,000 credits (roughly 10 minutes of speech, non-commercial), Starter is $6/month with 30,000 credits, Creator is $22/month with 121,000 credits, and plans scale up to $990/month for Business. Realtime pricing is billed separately on the usage model above — check the official pricing page for the current per-minute and per-token rates for these two specific models, since new-model pricing often moves in the first weeks after launch.

Where the money math matters: per-minute billing punishes long calls and chatty agents. If your agent holds 10-minute support conversations, the difference between Max and Mini compounds fast. Start with Mini for simple assistants and measure cost-per-resolved-call, not cost-per-minute — a cheaper model that resolves fewer calls can be the expensive choice.

The week's real question: Max, Mini, or Voxtral 3?

The Voxtral 3 launch deserves the direct comparison because it happened days apart and targets the same buyer: developers building voice agents. The two products are genuinely different architectures:

• Eleven Realtime is a speech-to-speech model: one black box that hears, reasons, and speaks. The tradeoff is control — you cannot swap, inspect, or independently configure the underlying components, and there is no text transcript (which also means text-based safety tooling does not apply directly).
• Voxtral 3 is a full stack: STT/TTS models plus a LiveKit-based voice-agent SDK in one streaming pipeline, 50+ languages, EU-hosted or self-deployable. More components to manage, but more to inspect and control — and for European customers, the data-residency story is the differentiator.

The voice-agent week did not stop there: it also overlapped with Grok Voice 3 and a steady drumbeat of voice-sales tooling across the industry. Voice is having its infrastructure moment, and the practical takeaway is boring but important: nobody's benchmarks are independent yet. Vendor charts are marketing until third parties reproduce them. The cheap test is the API-swap test — and ElevenLabs' OpenAI-compatible format exists precisely to make that test cheap.

Pros and cons

What is genuinely good:

• One model call replaces a multi-component pipeline, which should mean fewer integration seams and lower latency in theory.
• The OpenAI-compatible API lowers switching costs dramatically — side-by-side trials are a realistic afternoon's work.
• Mini's turn-taking handling addresses the most common "this feels like a robot" complaint.
• 90+ languages and a huge voice library make it usable for multilingual businesses without building a system per market.

What to watch out for:

• No independent benchmarks exist yet — every performance claim is the vendor's.
• No text transcript means no existing text-based moderation, logging, or compliance tooling.
• You cannot inspect or tune the internal components — when it fails, you debug the whole black box.
• One secondary report notes that WebRTC connections must move to WebSockets, which is an integration cost to check with your developer.
• HIPAA/Zero Retention only on Max — regulated-data users cannot use Mini.

Should you switch?

The measured answer: run a trial, don't migrate blind. The technical change is small (endpoint + model name), so set up a side-by-side test with real customer calls and compare the two things benchmarks never measure: does it resolve calls, and does it feel natural? Test Max on your hardest call scripts and Mini on your simplest ones. If your workload is support or multi-step tool use, Max is the obvious starting point. If it is a simple assistant and cost dominates, Mini. And if you handle regulated data, your choice is Max or nothing.

ElevenLabs started as a text-to-speech component company. With ElevenAgents, September's Eleven v4 TTS releases, and now Eleven Realtime, it is clearly moving up the stack into the full conversational platform business. This launch is less "new TTS model" and more "ElevenLabs wants to own your voice agent." Whether it earns that position depends on what happens when real production traffic hits it — and that part of the story has not been written yet.

---

Company claims about benchmarks and leaderboard positions come from ElevenLabs' October 20, 2026 launch materials. Pricing details are usage-based per ElevenLabs' pricing page — verify current rates before committing to a build.