ElevenLabs Eleven Realtime Review: 98.8% Speech Accuracy in 0.57 Seconds, but No Price List Yet

Building a voice agent has always meant wiring together three separate systems: a speech recognizer to hear the caller, a language model to think, and a speech synthesizer to talk back. On October 20, 2026, ElevenLabs launched Eleven Realtime, a pair of speech-to-speech models that collapse that pipeline into a single model. One component listens, reasons, and speaks — and ElevenLabs says the top variant answers spoken reasoning questions at 98.8% accuracy while starting to speak in 0.57 seconds.

Those are the vendor's numbers, and they deserve scrutiny. This review covers what Eleven Realtime actually is, where Mini and Max differ, what the benchmarks show, how the OpenAI-compatible API works, and the one big thing missing from the announcement: any pricing at all.

The short answer

• What it is: Two speech-to-speech models — Eleven Realtime v1 Mini and v1 Max — each a complete voice agent in one model: speech recognition, turn-taking, reasoning, and voice generation combined.
• Mini is for natural conversations at low cost — AI assistants, tutors, interactive characters. Max is for complex work — customer support, lead qualification, multi-step tool-calling workflows.
• Performance claims (vendor-reported, Artificial Analysis-evaluated): Max answers 98.8% of Big Bench Audio reasoning questions vs 75.9% for Mini; Max starts speaking in 0.57 seconds, Mini in 0.62. Mini wins conversational dynamics at 95.8%.
• Access: Through ElevenAgents (phone, web, WhatsApp, contact-center channels) or a Realtime API matching OpenAI's Realtime API spec — change the endpoint and model name to migrate.
• Pricing: Not published at launch. In ElevenAgents you see an estimated cost per minute in the model selector; the API bills per minute of connected audio plus tokens, with the per-token price still undisclosed at launch. Verify before you build.

What Eleven Realtime actually changes

The typical voice agent is plumbing: connect speech-to-text to a language model, pipe the text into a speech synthesizer, then tune the pauses so the caller doesn't hear awkward silence or get talked over. Each handoff adds latency and another place to debug. ElevenLabs' pitch is blunt: "You shouldn't have to become a configuration expert to build a voice agent that sounds natural."

Eleven Realtime replaces the pipeline with one end-to-end model that handles speech recognition, voice activity detection, turn-taking, reasoning, and voice generation internally. The practical consequences:

• Lower latency — nothing is serialized through three services. Max averages 0.57 seconds to first audio.
• Better turn-taking — the model hears prosody directly instead of reading a transcript, so it can tell when a caller has finished versus when they're just saying "mm-hmm," according to ElevenLabs. Turn-taking on real call audio is one of the hardest details to get right in a pipeline.
• Less control — as AlexTech noted in its launch analysis, a pipeline lets you swap one component or intervene on the intermediate transcript; a single model gives you fewer handles. If you want to inspect or edit the exact transcript the agent heard, a monolithic model leaves you less room to do it.

Both models support more than 90 languages and any of the 17,000+ voices in the ElevenLabs library, so localization becomes a voice choice rather than an engineering project.

Mini vs Max: which one, and when

* Compliance details per launch coverage; confirm in ElevenLabs' current docs before signing anything regulated.

The most honest line in the launch materials might be the recommendation order: start with Mini and move to Max only when the agent needs to follow long instructions, chain tool calls, or meet stricter compliance rules. The benchmark split backs that up — Max is 23 percentage points ahead on reasoning and 27 points ahead on tool use, but Mini is cheaper and measurably better at the pauses, interruptions, and backchannels that make a conversation feel natural.

The benchmarks, with the fine print

Every number in this section comes from ElevenLabs' launch announcement. The company says Artificial Analysis ran the evaluations independently — meaningful, but these are still company-published launch-day figures with no independent reproduction yet. Read them as the vendor's best-case demonstration, not a settled verdict.

Big Bench Audio tests reasoning through spoken questions and answers: the model hears a question and is scored on its spoken answer. Max's 98.8% over 1,000 questions is genuinely striking; Mini's 75.9% sits in line with the broader field. The gap tells you these two models are in different capability classes, not just different price classes.

Speed: Max starts speaking in 0.57 seconds on average, Mini in 0.62 seconds. ElevenLabs claims no model on the Artificial Analysis speech-to-speech leaderboard is currently both faster and more accurate. That claim holds only against the leaderboard's current entries — sub-second first-audio is table stakes for any serious voice agent in 2026.

Full-Duplex-Bench measures conversational dynamics — pauses, turn-taking, interruptions, backchannels. Mini's 95.8% beats Max's 94.1%. It's a small gap, but it matters: the model optimized for reasoning isn't the model optimized for sounding human.

Speech Agent Arena scores blind live conversations by human preference and task success: Max 998, Mini 988 — a slim lead for Max. On tool-use tasks, Max completes 87.1% vs 60.3% for Mini — the clearest reason to pay more: if your agent looks things up or walks a multi-step process, Max is in a different league.

The pricing problem: ElevenLabs published none of it

Here is the uncomfortable part of an otherwise strong launch. The announcement contains no pricing, no per-minute rates, no per-token figures, no credit math. ElevenLabs describes Mini only qualitatively as "low cost."

What launch coverage has established so far:

• In ElevenAgents, the Realtime models follow standard per-minute pricing, and the model selector displays an estimated cost per minute before you commit.
• Through the Realtime API, billing is per minute of connected audio plus tokens, with the token price the same regardless of plan. The actual per-token price was not announced at launch — check elevenlabs.io/pricing for the live numbers before architecting anything.

For context on the bill: voice-agent costs stack up from three directions — the minutes the agent is connected, the tokens the reasoning consumes, and whatever the platform charges. A model that "costs less per minute" can still be the expensive choice if it needs longer conversations to finish the same task. Until ElevenLabs publishes both axes of the rate card, cost comparisons with rival offerings are guesswork.

How to get it: ElevenAgents vs the Realtime API

ElevenAgents (recommended path). Pick a Realtime model from the dropdown and deploy across phone, web, WhatsApp, and contact-center platforms. This route includes workflows, a knowledge base, guardrails, testing, and monitoring — the operational tooling a production voice agent lives or dies by.

Realtime API. This is the developer play: the API follows OpenAI's Realtime API specification, so teams already running on OpenAI's Realtime API can migrate by changing the endpoint and the model name — the OpenAI SDK works as-is, including with packages like @openai/agents-realtime. Three catches to know before you migrate:

• Connections run over WebSocket only (at wss://api.elevenlabs.io/v1/convai/realtime). If your current integration uses WebRTC, you have to move it to WebSocket first.
• The Realtime API runs separately from ElevenAgents — no channels, no knowledge base; you bring your own tooling. Example model names look like eleven_realtime_v1_mini — expect these to evolve; pin them from the docs at build time.

The crowded field

Eleven Realtime lands in a market that got busy the same week. Mistral launched Voxtral 3 — a voice stack with transcription, speech synthesis, and a LiveKit-based SDK that can run on the customer's own infrastructure — within days of this launch, while OpenAI and Google compete in realtime voice too.

ElevenLabs' bet is simplification: one model instead of a pipeline, and an OpenAI-compatible API that makes switching cheap to try. The bet looks aimed squarely at enterprise, where large organizations now account for 55% of ElevenLabs' revenue, up from 40% twelve months ago. Voice agents are moving from demos to production call centers, and the vendor that removes the plumbing wins the boring, lucrative middle of the market.

Pros and cons

What works

• Real single-model voice agents — no pipeline assembly, and the "mm-hmm" detection addresses a genuinely hard conversational problem.
• Strong reasoning numbers for Max — 98.8% on Big Bench Audio is a headline figure, with Artificial Analysis backing the methodology.
• Honest product split — Mini wins the dynamics benchmark, and ElevenLabs tells most users to start there instead of upselling everyone to Max.
• Zero-friction migration — OpenAI Realtime API compatibility means existing voice-agent teams can test it with an endpoint and model-name change.
• Scale-ready surface — 90+ languages and 17,000+ voices, plus ElevenAgents' phone/WhatsApp/contact-center deployment and guardrails.

What to watch

• No published pricing — the biggest gap in the launch. You cannot cost a project from the announcement.
• Company-reported benchmarks — independently run by Artificial Analysis, but still launch-day figures with no independent reproduction yet.
• Less control than a pipeline — you trade per-component control and transcript-level intervention for simplicity.
• WebSocket-only API — WebRTC integrations must be reworked to connect.
• Tool-use gap between variants — Mini's 60.3% tool completion means the cheap option may not fit agentic workflows.

FAQ

What is ElevenLabs Eleven Realtime?

A family of two speech-to-speech models — v1 Mini and v1 Max, launched October 20, 2026 — that each work as a complete voice agent. One model handles listening, reasoning, turn-taking, and speaking instead of chaining three separate systems.

How much does Eleven Realtime cost?

ElevenLabs did not publish per-minute or per-token pricing at launch. In ElevenAgents, the model selector shows an estimated cost per minute; through the Realtime API, you pay per minute of connected audio plus tokens, with the per-token price undisclosed at launch. Check the live pricing page — launch announcements are not rate cards.

Should I use Mini or Max?

Start with Mini for natural conversation — AI assistants, tutors, interactive characters — where it is cheaper and actually scores higher on conversational dynamics (95.8%). Move to Max when the agent needs strong reasoning, heavy tool calling (87.1% vs 60.3% task completion), or compliance options like Zero Retention Mode and HIPAA.

Is Eleven Realtime compatible with OpenAI's Realtime API?

Yes. The Realtime API follows OpenAI's Realtime API specification — the OpenAI SDK works as-is; you change the endpoint and the model name. One caveat: it connects over WebSocket only, so WebRTC-based integrations must migrate first.

What languages and voices does it support?

More than 90 languages and any of the 17,000+ voices in the ElevenLabs library, per the launch announcement.

Key takeaways

• ElevenLabs' Eleven Realtime collapses the voice-agent pipeline — speech recognition, reasoning, turn-taking, voice generation — into a single model, in two variants: v1 Mini for natural conversation and v1 Max for complex tool-calling work.
• The benchmark split is unusually honest: Max leads reasoning (98.8%) and tool use (87.1%), but the cheaper Mini wins conversational dynamics (95.8%), and ElevenLabs recommends starting there.
• All headline figures are company-reported from launch day — Artificial Analysis ran the evaluations, but no independent reproduction exists yet.
• The Realtime API is OpenAI-compatible with WebSocket-only connections, making migration from OpenAI's Realtime API cheap to try but requiring a WebRTC rework for some integrations.
• The launch's biggest hole is pricing: no per-minute or per-token rates were published, so verify the live numbers in ElevenAgents' model selector or the pricing page before you build.

Published analysis as of October 22, 2026. Voice-agent pricing and model availability move fast — confirm the current rate card at elevenlabs.io/pricing before making decisions.