# Eleven Realtime v1 Mini and Max: API Pricing, Benchmarks and Limits

ElevenLabs packages a complete voice pipeline into two model choices, but API compatibility, billing and regional support need a closer look.

By Paras Madan | 2026-10-21


ElevenLabs launched Eleven Realtime v1 Mini and Max on October 20, 2026. The two voice-agent models combine speech recognition, turn-taking, reasoning and speech output into a managed system, available through ElevenAgents and an OpenAI-compatible Realtime API. Developers get fewer components to configure, with different tradeoffs for conversational and tool-heavy work.

## What Eleven Realtime changes

The [launch announcement](https://elevenlabs.io/blog/realtime) positions Mini for assistants, tutors and interactive characters, and Max for complex instructions and multi-step tasks. Both support more than 90 languages and the company's library of over 17,000 voices.

This changes the integration boundary. Instead of selecting and tuning every part of the audio pipeline, a team chooses a complete configuration. It still needs to define the agent's instructions and tools, then test whether the conversation completes the intended task.

The architecture matters here. ElevenLabs' migration documentation says these are **cascaded systems, not fused models**. Separate components are tuned to work together, and their individual model identities are not public. A speech-to-speech interface does not, by itself, tell you how a system reasons internally.

## Eleven Realtime API pricing

The [public API pricing page](https://elevenlabs.io/pricing/api) lists a connected-audio charge plus model token usage:

| Charge | v1 Mini | v1 Max |
| --- | --- | --- |
| Additional connected audio | $0.08 per minute | $0.08 per minute |
| Input tokens | $0.20 per million | $1.80 per million |
| Output tokens | $6 per million | $9 per million |
| Connected audio above concurrency limit | $0.16 per minute | $0.16 per minute |

These are USD prices before taxes. Included minutes are estimates based on typical token usage, so conversation length and tool use affect how far a plan goes.

For example, 1,000 additional connected minutes costs $80 in the audio component at the standard rate. That is not the complete invoice: input and output tokens add to it, and burst usage changes the audio rate. A short, direct conversation and a long troubleshooting call can therefore have different economics even at the same duration.

## Mini versus Max benchmarks

ElevenLabs reports independently conducted Artificial Analysis evaluations. Its published results give Max 98.8% on Big Bench Audio against Mini's 75.9%, with average time to first audio of 0.57 and 0.62 seconds respectively. On Full-Duplex-Bench conversational dynamics, Mini scores 95.8% and Max 94.1%.

Those results suggest different strengths rather than a universal winner. Mini's conversational behaviour can matter more than difficult reasoning for a character or simple assistant, while Max's stronger reasoning may justify its higher token price for a support workflow.

The [Artificial Analysis methodology](https://artificialanalysis.ai/speech-to-speech) explains that Big Bench Audio contains 1,000 spoken reasoning questions across four categories. Time to first audio measures when speech begins, not how long an entire business transaction takes. Treat these as evaluation results, not a latency guarantee for your network, tools or callers.

## OpenAI Realtime API migration

The [compatibility guide](https://elevenlabs.io/docs/eleven-agents/guides/openai-realtime-compatibility) describes reusing OpenAI's Realtime SDK and event protocol over WebSocket. The endpoint is `wss://api.elevenlabs.io/v1/convai/realtime`, with model IDs `eleven_realtime_v1_mini` and `eleven_realtime_v1_max`.

Your backend creates a short-lived client secret, then the browser uses that credential to connect. The permanent ElevenLabs API key stays on the backend.

**WebSocket compatibility is not WebRTC compatibility.** Existing WebRTC applications need to change transport. The guide also says the SDK only wires audio automatically for WebRTC, so a WebSocket client must capture and play PCM16 audio at 24kHz itself. Server-side OpenAI Realtime SDK integration documentation is not yet published in this guide.

The direct Realtime API is separate from ElevenAgents. Choose the platform when you need its channels, knowledge base, workflows and operational tools rather than assuming they come with the compatible socket endpoint.

## Availability and deployment restrictions

The [Realtime model documentation](https://elevenlabs.io/docs/eleven-agents/customization/realtime-models) says workspace access is rolling out gradually. If the Voice mode control is missing, the API also rejects the Realtime model setting for that workspace.

Mini is available in the US, EU, India and Singapore environments, but not under Zero Retention Mode or HIPAA settings. Max supports those settings in the US and EU, but is unavailable in the India and Singapore data-residency environments. Max also requires a token limit of at least 1,024, or no limit.

Custom-mode agents retain value when a team needs a particular LLM, custom model or component-level control. Transfers between Realtime and Custom agents are unsupported, so migration should include the call-routing graph rather than only the first agent a caller reaches.

The useful test is a complete conversation: interruptions, a failed lookup, a correction and the final action. Packaging the voice stack reduces configuration work. It makes careful workflow evaluation easier to start, not unnecessary.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn AI launches and product updates into clear stories that developers and founders want to read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).