Key specifications
| Specification | Realtime v1 Max | Realtime v1 Mini |
| Big Bench Audio | 98.8% | 75.9% |
| Time to first audio | 0.57 s | 0.62 s |
| Full-Duplex-Bench | 94.1% | 95.8% |
| Tool-use success | 87.1% | 60.3% |
| Languages | 90+ | 90+ |
| Voices | 17,000+ | 17,000+ |
| API | WebSocket | WebSocket |
| ElevenAgents | Yes | Yes |
Max is optimized for reasoning and tool use. Mini targets cheaper, high-volume conversational workloads.
Realtime v1 Max starts speaking in about 0.57 seconds.
Max vs Mini
Max:
- 98.8% on Big Bench Audio
- 87.1% tool-use success
- 570 ms time to first audio
- better suited for support agents, workflows and complex instructions
Mini:
- 75.9% on Big Bench Audio
- 60.3% tool-use success
- 620 ms time to first audio
- slightly better Full-Duplex-Bench result: 95.8%
- designed for lower-cost natural conversations
Max leads Mini by 22.9 percentage points on Big Bench Audio and 26.8 points on tool-use tasks.
Architecture
Eleven Realtime is not a single end-to-end multimodal model.
It uses a cascaded architecture combining:
speech recognition → language model → speech generation
ElevenLabs manages and optimizes these components as one realtime system.
API details
The models are available through ElevenAgents and a WebSocket Realtime API.ElevenLabs also provides compatibility with the OpenAI Realtime API protocol, reducing migration work for existing integrations.
Example audio configuration uses:
- PCM16
- 24 kHz audio
- WebSocket streaming
Compared with the previous ElevenLabs setup
There is no direct “Realtime v0.” Realtime v1 replaces the previous agent architecture where developers typically configured separate ASR, LLM and TTS components.
| Previous setup | Eleven Realtime v1 |
| Separate ASR | Integrated |
| Separate LLM | Integrated |
| Separate TTS | Integrated |
| Manual component tuning | Pre-optimized stack |
| More model-level control | Simpler integration |
| Higher configuration complexity | Single realtime interface |
The main trade-off is flexibility: developers get lower integration complexity, but less control over the underlying models.
Bottom line
Realtime v1 Max is the stronger model for reasoning and tool-driven agents, with 98.8% Big Bench Audio accuracy and 570 ms response latency.
Realtime v1 Mini sacrifices reasoning performance for lower-cost conversational use while maintaining similar latency at 620 ms.
The main product change is not only speed: ElevenLabs is packaging ASR, reasoning and TTS into one optimized realtime voice-agent stack.
