ElevenLabs has unveiled Eleven Realtime, a new collection of specialized speech-to-speech artificial intelligence models designed to power natural voice conversations. The flagship model variant, Eleven Realtime Max, registered a 98.8 percent benchmark score on Big Bench Audio while maintaining a latency response time of 0.57 seconds. The company claims this performance combination establishes a top benchmark for simultaneous speed and accuracy among publicly ranked voice models.

The release includes two distinct model architectures optimized for different operational needs: the low-cost v1 Mini, engineered to recognize conversational nuances such as backchannel pauses, and the v1 Max, designed for handling intricate enterprise workflows like automated customer support. Both voice models are accessible through the company’s proprietary ElevenAgents software platform as well as an OpenAI-compatible Realtime API.

Commercial access is structured via a combined per-minute usage rate and standard processing token charges, with the Max variant commanding a higher price point. The launch underscores growing market competition to deliver responsive, human-like voice infrastructure to enterprise customer service and consumer assistant applications.

Why it matters

  • Product teams can deploy lower-latency voice agents capable of handling complex conversational interruptions and enterprise workflows.

  • Developers gain direct API access to high-accuracy speech models compatible with existing OpenAI endpoint structures.

Source: the-decoder.com