ElevenLabs launches Realtime Mini and Max through an OpenAI-compatible API

• Eleven Realtime ships two voice-agent configurations: Mini for natural conversation and Max for complex tasks, available through ElevenAgents and an API in ElevenLabs' announcement.
• Max scores 98.8% on Big Bench Audio and starts speaking in 570 ms, according to ElevenLabs' benchmark report.
• Existing OpenAI Realtime integrations can retain the SDK and event protocol, but the migration example uses WebSocket. WebRTC is unsupported at launch.

Mini and Max use a cascade of distinct component models, a more interesting detail than the speech-to-speech branding. The browser example mints ten-minute credentials, while Mini outperforms Max on conversational dynamics.

ElevenLabs positions the two variants around different workloads:

• v1 Mini: lower-cost assistants, tutors, and interactive characters. Its launch description emphasizes detecting completed turns, recovering from interruptions, and continuing through backchannels such as “mm-hmm.”
• v1 Max: customer support, lead qualification, and multi-step workflows involving reasoning and tool calls.

Both support more than 90 languages and the company's library of 17,000-plus voices, according to the October 20 launch post.

The packaged stack contains five components listed in the developer announcement:

1. Speech recognition.
2. Voice activity detection.
3. Turn-taking.
4. A language model.
5. Text-to-speech.

Each component is optimized to work within the combined system. ElevenLabs has not disclosed the identities of the underlying models, according to its architecture FAQ.

Artificial Analysis independently evaluated both models across three public benchmarks, according to the launch report:

• Big Bench Audio: 1,000 spoken reasoning questions, scored on the spoken answer.
• Full-Duplex-Bench: pauses, turn-taking, interruptions, and backchannels.
• Speech Agent Arena: blind live comparisons for human preference, plus task-success results.

| Metric | Mini | Max |
| --- | ---: | ---: |
| Big Bench Audio accuracy | 75.9% | 98.8% |
| Time to first audio | 620 ms | 570 ms |
| Tool-use task success | 60.3% | 87.1% |
| Full-Duplex-Bench | 95.8% | 94.1% |
| Speech Agent Arena preference Elo | 988 | 998 |

The “no model is both faster and more accurate” claim describes a speed-accuracy frontier. The chart's footnote identifies a higher-scoring model outside its three-second plotting range: StepAudio 3 Realtime reaches 99.7% at 8.83 seconds.

The chart in the later developer thread labels ElevenLabs' measurements as medians and competitors' measurements as leaderboard averages. The launch post describes time to first audio as an average.

ElevenLabs promises an endpoint-and-model swap in its API announcement. The compatibility guide retains @openai/agents-realtime and the same client and server events.

Browser integrations have additional plumbing described in the migration guide:

• Transport: WebSocket only at launch. Existing WebRTC integrations require a transport change.
• Audio: the application captures and plays PCM16 audio at 24 kHz. The OpenAI SDK only wires audio automatically for its WebRTC transport.
• Authentication: a backend mints an ElevenLabs client secret scoped to a session configuration; the browser passes that secret to the OpenAI SDK.
• Configuration: the model string selects behavior directly, rather than referencing an ElevenAgents dashboard agent_id.

ElevenLabs also offers a migration skill for coding agents:

Selecting Mini or Max in ElevenAgents preserves existing agent configuration, according to the product FAQ. The platform provides:

• Channels: phone, web, WhatsApp, and contact-center platforms.
• Operations: knowledge bases, guardrails, testing, and monitoring.

The standalone Realtime API runs separately from ElevenAgents and does not expose these platform features, the FAQ states.

The API pricing page lists both Mini and Max at $0.08 per minute plus LLM usage. That makes the base charge $4.80 per connected hour before token costs.

The API meters connected audio time and adds token charges, with token prices identical across subscription plans, according to the pricing FAQ. Inside ElevenAgents, billing uses standard platform per-minute pricing plus LLM token costs; the model selector displays an estimated cost per minute.