ElevenLabs offers two integrated voice-agent models with different task roles

One model choice replaces several configuration choices

ElevenLabs is offering speech-to-speech models that combine the elements of a voice agent rather than asking builders to assemble each audio component separately. In its October 20 announcement, the company positions Realtime v1 Mini for conversational assistants and characters, and v1 Max for tasks such as customer support and multi-step processes.

Both are offered within ElevenAgents and through a separate Realtime API. The company describes compatibility with OpenAI’s Realtime interface, but the launch guidance identifies a practical restriction: this API uses WebSocket transport. Existing WebRTC applications need a transport change.

Check the product boundary

The standalone API does not automatically include the channels, knowledge base and other features provided by ElevenAgents. Pricing and the appropriate account configuration should be checked for the intended deployment. Published speed and task-performance figures are vendor-presented evidence; CGH has not replicated those results.

CGH perspective

The choice should follow the conversation’s hardest moment. A tutor needs to handle a hesitant answer; a service agent may need to correct an order without repeating an action. A useful trial would compare interruptions, recovery from a failed tool call and the ability to hand a conversation to a person. Measure completed tasks and understandable responses, not voice quality alone. An integrated model can simplify construction while leaving those acceptance decisions with the team deploying it.