ElevenLabs launched Eleven Realtime on October 20, a new family of speech-to-speech models built for voice agents. The key point is that each model is a complete voice agent: developers no longer need to wire together speech recognition, a language model and speech synthesis themselves. Teams building phone assistants or voice features on websites can build an agent with a single component.
The family has two variants. v1 Mini is designed for natural conversations at low cost, while v1 Max targets more complex work such as customer support and multi-step tool use. Both cover more than 90 languages and over 17,000 voices. According to ElevenLabs, Mini can tell when someone has finished speaking or is just saying "mm-hmm".
Image: ElevenLabs
On performance, the standout figure concerns Max: in the Big Bench Audio spoken reasoning test, evaluated by Artificial Analysis on 1,000 questions, it answers 98.8% correctly, against 75.9% for Mini. Audio starts on average after 0.57 seconds for Max and 0.62 seconds for Mini. ElevenLabs claims no model on the Artificial Analysis speech-to-speech leaderboard is both faster and more accurate. These are company-reported figures.
For developers, the models can be used in two ways: through the ElevenAgents platform or through a Realtime API compatible with OpenAI's protocol, which reduces the work for teams already integrated with that format. Pricing is per minute plus token costs, with Max costing more than Mini.
The launch arrives in a crowded market. In early October Mistral presented Voxtral 3, a voice stack with transcription, speech synthesis and a LiveKit-based SDK that can run on the customer's own infrastructure. OpenAI and Google also compete in this area. ElevenLabs' move is to simplify the architecture, with one model instead of a three-component pipeline.
Image: ElevenLabs
What changes for developers
Choosing a single model over a modular pipeline is not neutral. With a pipeline you can replace one piece, such as the speech synthesizer or the language model, and control each stage more closely. A speech-to-speech model reduces latency and complexity, but leaves less control over individual components: those who want to intervene on the intermediate transcript or change the voice mid-flow have fewer options.
No comments yet. Be the first!