SpaceXAI has released Grok Voice 3.0, an updated realtime speech-to-speech AI model that integrates reasoning, tool calls, and low-latency voice output. The model achieved first place on the Artificial Analysis Speech-to-Speech Index with a score of 86.9, outperforming competitors including OpenAI’s GPT-Live-1 and Google’s Gemini 3.8 Live.

Exposed via the grok-voice-3.0 WebSocket endpoint, the API maintains compatibility with the OpenAI Realtime API and keeps pricing at $0.08 per minute. According to benchmark metrics, Grok Voice 3.0 recorded a 0.81-second time to first audio—40% faster latency than GPT-Live-1—while supporting server-side voice activity detection and in-session tool execution across search platforms and remote MCP servers.

Although parameter details were not disclosed, reasoning is enabled by default for voice interactions. The updated API allows developers to configure reasoning effort levels, while Grok Voice 2.0 has been deprecated as existing integrations automatically migrate to the new model endpoint.

Why it matters

  • Sets a new technical benchmark for realtime speech-to-speech agents with sub-second latency and direct tool execution.

  • Maintains drop-in API compatibility with OpenAI’s Realtime protocol, lowering developer migration barriers.

  • Highlights intense competition among top AI labs in pricing and performance for interactive voice models.

Source: marktechpost.com