xAI Launches Grok Voice 3.0, Its Fastest Voice Model Yet

xAI, Elon Musk's AI company, released Grok Voice 3.0, a new model for real-time voice conversations. Independent benchmark firm Artificial Analysis ranked it #1 among voice models, with 86.9 points versus 81.3 for the previous version. It replies in under a second, 40% faster than OpenAI's GPT-Live-1, at similar accuracy. The model already runs in Tesla, Musk's car company, and in the call center of Starlink, his satellite internet provider. Starlink's call center handles 20,000 customer calls a day, and about 70% get resolved without a human.

For voice support agents, this is not just another benchmark score. Twenty thousand calls a day, with seventy percent resolved automatically, is already a real production workload, not a demo. The Big Bench Audio test measures what it costs a voice model to finish a task. There, Grok Voice 3.0 comes in cheaper than GPT-Live-1 and GPT-Realtime-2.1. That means an agent handling thousands of calls a day should also run cheaper. It is a reasonable pick for the voice layer of a support agent. The reasoning behind that agent is still better left to models built for reasoning. Claude Opus and Fable, made by Anthropic, remain the stronger choice there.