xAI Grok Voice 3.0

xAI Grok Voice 3.0 is a native speech-to-speech model released in October 2026, delivering full-duplex real-time conversations over WebSockets with built-in reasoning capabilities. It scored 97 percent on the Full Duplex Bench evaluation, the highest published turn-taking and interruption score in the category, and 71.6 percent on the tau-Voice agentic benchmark, up from 56.5 percent in Grok Voice 2.0. Grok Voice 3.0 supports 25-plus languages with 38-plus language STT coverage, voice cloning from short reference clips up to 120 seconds, and an extensive tool-calling framework that orchestrates multi-step external API calls mid-conversation. At $0.08 per minute flat with no additional token costs, the pricing is predictable. SDK examples cover Python, JavaScript, Twilio, WebRTC, and iOS for rapid integration. Grok Voice 3.0 targets developers building multilingual voice products who need a single native model rather than a cascade pipeline. Its deployment in Tesla and Starlink, where it resolves approximately 70 percent of calls without human handoff, demonstrates enterprise-scale reliability. The built-in reasoning layer handles complex multi-step requests that would require LLM orchestration in a cascade architecture.