Voice Generation
Eleven Realtime Mini vs Max: Which ElevenLabs Model for Voice Agents
Eleven Realtime v1 Mini vs Max compared on latency, spoken reasoning and tool use, with the use cases each model fits and the pricing caveats before building a voice agent.

Some links are partner links: if you subscribe through them, we may earn a commission, at no extra cost to you. The crowd verdicts stay independent.
Eleven Realtime Mini vs Max is the first real decision a team faces when building a voice agent on ElevenLabs' new speech-to-speech API: Max is faster and far more accurate, Mini is built for cheaper, natural conversation. The independent benchmarks published with the October 2026 launch make the trade-off unusually clear, and the price per minute is the one number still missing.
Max for tasks, Mini for talk
Pick Eleven Realtime v1 Max for support, lead qualification and multi-step tool calls: it scores 98.8% on Big Bench Audio and 87.1% on tool use, against 75.9% and 60.3% for Mini. Pick Mini for assistants, tutors and characters, where it actually wins on conversational flow (95.8% Full-Duplex-Bench). ElevenLabs plans start at $5/month.
What Eleven Realtime is (and what changed on October 20)
Eleven Realtime is a family of speech-to-speech models: audio goes in, audio comes out, behind a single model ID. Until now, an ElevenLabs voice agent chained three steps (speech-to-text, a language model, then text-to-speech). A native speech-to-speech model removes two hand-offs, which is where most conversational lag used to come from.
Two models shipped on October 20, 2026 (the official announcement was updated on October 29):
- Eleven Realtime v1 Mini: positioned for natural conversation at low cost (AI assistants, teachers, interactive characters).
- Eleven Realtime v1 Max: positioned for complex tasks with heavy tool calling (customer support, lead qualification, multi-step workflows).
Both speak 90+ languages and can use any of the 17,000+ voices in the ElevenLabs library. That catalogue figure matters: the voice a caller hears is the same voice a brand may already use for its YouTube narration or audiobooks, a consistency no OpenAI-only stack offers.
Three weeks later, at its New York Summit on November 11, ElevenLabs added Memory, Routines and a preview of an Interaction Engine to its agents platform, plus Scribe v3 Medical for transcription. Realtime is the voice layer under that whole stack.
Eleven Realtime Mini vs Max: latency and accuracy benchmarks
The scores below come from Artificial Analysis, an independent evaluator, and were published by ElevenLabs with the launch.
| Benchmark | What it measures | v1 Mini | v1 Max | Winner |
|---|---|---|---|---|
| Time to first audio | Average delay before the agent starts speaking | 0.62 s | 0.57 s | Max |
| Big Bench Audio | Spoken reasoning, 1,000 questions | 75.9% | 98.8% | Max |
| Tool use tasks | Calling functions correctly mid-conversation | 60.3% | 87.1% | Max |
| Full-Duplex-Bench | Conversational dynamics (turn-taking, interruptions) | 95.8% | 94.1% | Mini |
| Speech Agent Arena | Human preference score | 988 | 998 | Max |
Three readings stand out.
Latency is not the differentiator. A 0.05-second gap (0.62 s vs 0.57 s) is below what a caller can perceive. Both models sit comfortably under the one-second threshold where a phone conversation starts to feel broken. Choosing Mini "to be faster" makes no sense: Max is the faster of the two.
Accuracy is the real split. The Big Bench Audio gap is 22.9 points, and the tool use gap is 26.8 points. In practice, a 60.3% tool use score means roughly four calls in ten go wrong: a booking lands on the wrong date, an order lookup fails, a CRM field stays empty. For any agent whose job is to do something, Mini is not production-ready on these numbers.
Mini talks slightly better. On Full-Duplex-Bench, which tests how naturally a model handles interruptions, pauses and overlapping speech, Mini edges Max (95.8% vs 94.1%). For a language tutor or a game character, where the conversation itself is the product and no API gets called, that is the metric that matters.
Which model to choose by use case
The verdict of the arena follows the benchmarks, not the marketing tiers.
| Use case | Recommended model | Why |
|---|---|---|
| Call triage and routing | Mini | Few tool calls, short exchanges, cost matters at volume |
| Language tutors, companions, game characters | Mini | Best turn-taking score (95.8%), no tools involved |
| Customer support with order or account lookups | Max | 87.1% tool use vs 60.3% |
| Lead qualification feeding a CRM | Max | Multi-step logic plus writes to external systems |
| Appointment booking and multi-step workflows | Max | Every failed call is a lost booking |
A practical pattern for contact centres: route the first seconds of a call through a cheaper triage agent, then hand off to a Max agent once the caller's intent requires actions. Whether that split actually saves money depends on the per-minute price, which brings up the main caveat.
ElevenLabs
The realism benchmark for AI voices: best-in-class TTS and cloning at a premium price
Partner link. The crowd verdicts stay independent.
Eleven Realtime API: integration, compatibility and pricing caveats
The integration story is the strongest part of the launch. The Realtime API connects over WebSocket and follows OpenAI's Realtime specification, so the OpenAI SDK works unchanged: an existing OpenAI Realtime integration switches by changing the endpoint and the model name. For teams already running a voice agent on OpenAI, testing ElevenLabs costs an afternoon, not a rewrite.
Three limits deserve attention before committing:
- No direct WebRTC. Browser or mobile integrations built on WebRTC must move to WebSocket first. That is real work for client-side apps.
- The API is separate from ElevenAgents. Phone channels, the knowledge base and the other platform features require ElevenAgents, which ElevenLabs recommends as the default entry point.
- No published per-minute price. The launch post states no Realtime pricing. Any cost comparison with OpenAI or a cascaded pipeline is guesswork until ElevenLabs publishes it, and no figure is quoted here for that reason.
What is known is the subscription ladder, verified on the ElevenLabs pricing page in October 2026: Starter at $5/month (annual), Creator at $22, Pro at $99, Scale at $299 and Business at $990. The free tier gives 10,000 credits per month for testing. The catalogue's main warning applies here too: credits burn fast, regenerations still consume them, and users report real costs up to about 2.8x the advertised per-character rate. Long agent sessions on a high-volume line will stress that system harder than narration ever did.
How ElevenLabs compares for voice agents
On GLAD-AI-TOR, ElevenLabs holds an 81% crowd score from 12 votes (11 recommend, 1 avoid), the highest in the voice arena, and "Voice agents / IVR" is one of its four best-for use cases. Two other catalogue tools are often shortlisted for the same job:
| Tool | From | Crowd score | Voice agent angle |
|---|---|---|---|
| ElevenLabs | $5/mo | 81% (12 votes) | Native speech-to-speech (Realtime), 17,000+ voices, OpenAI-compatible API |
| Resemble AI | $2/mo per voice clone | 50% (no votes yet) | Streaming TTS at ~200 ms, deepfake detection, on-prem; not an end-to-end agent platform |
| WellSaid Labs | $10/mo (annual) | 50% (no votes yet) | Licensed actor voices, IVR via enterprise API; API pricing not published |
Resemble AI is the alternative for regulated industries: metered TTS at $0.0005 per second (about $1.80 per audio hour), SOC 2 Type II, an on-premise option and built-in deepfake detection. It remains a voice layer, though: conversation logic needs separate tooling. WellSaid Labs is a brand-safe narration studio first; its voice agent use goes through a sales-gated enterprise API. Neither offers a native speech-to-speech model today. The full field is ranked on the ElevenLabs alternatives page, and the head-to-head is on ElevenLabs vs Resemble AI.
Teams that prefer a cascaded pipeline (their own transcription, a cheap LLM, then TTS) keep full control over the "brain" of the agent. That control comes at the cost of latency, which is exactly what Realtime removes.
ElevenLabs
The realism benchmark for AI voices: best-in-class TTS and cloning at a premium price
Partner link. The crowd verdicts stay independent.
The verdict
- Default to Max for anything that calls a tool. 87.1% tool use and 98.8% Big Bench Audio against 60.3% and 75.9% for Mini: a 26.8-point gap on actions is too wide for support or sales.
- Use Mini when the conversation is the product. Tutors, companions and characters benefit from its 95.8% Full-Duplex-Bench score, and no tool calls means its weak spot never shows.
- Do not pick Mini for speed. Max is faster (0.57 s vs 0.62 s), and both are well under one second.
- Budget with caution. No Realtime per-minute price is published yet; plan against the subscription ladder ($5 to $990/month) and the documented credit burn, then measure real usage on a pilot.
- Migration from OpenAI Realtime is cheap, unless the stack depends on WebRTC.
The crowd has already ranked the provider: 81% recommend ElevenLabs. To compare it with every other voice tool on price and score, see the best AI voice generators ranking, and read how ElevenLabs credits really work before sizing a plan.
Keep exploring
Every claim above is backed by the arena's live data: crowd votes, verified pricing, honest pros & cons.
More from the arena journal

Voice Generation
HeyGen Voice Turbo Tops the TTS Leaderboard: Free Until November 30, Then What?
HeyGen Voice Turbo is free until November 30, 2026, then $4 per 1M characters. What its #1 cloned-voice ranking proves, and how it compares with ElevenLabs v4 on cost.
Nov 2, 2026 · 7 min read

Voice Generation
ElevenLabs Eleven v4 Launches: The Most Emotive AI Voice Model Yet
ElevenLabs releases Eleven v4, their most emotionally expressive AI voice model. Here is what changed, what it costs, and whether it is worth the upgrade.
Oct 5, 2026 · 3 min read

Voice Generation
ElevenLabs CLI v1: What Developers Can Do From the Terminal
ElevenLabs released CLI v1 on August 24, 2026. Here is what developers can actually build with it: batch TTS, agents-as-code, CI/CD pipelines, and machine-readable API calls.
Sep 9, 2026 · 6 min read