xAI just dropped its flagship voice agent model, grok-voice-think-fast-1.0, and it's already available via API. Designed for complex, multi-step workflows like customer service and telemarketing, the model can reason in the background without adding response latency and natively supports over 25 languages.
On the τ-voice Bench — a full-duplex voice agent evaluation developed by Sierra that simulates real-world conditions like background noise, accents, and interruptions — grok-voice-think-fast-1.0 scored 67.3%, blowing past the competition. Second place went to Google's gemini-3.1-flash-live-preview-thinking-high at 43.8%, followed by OpenAI's gpt-realtime-1.5 at 35.3%. Even xAI's own previous generation, grok-voice-fast-1.0, managed only 38.3%. The new model topped the charts across all three sub-domains: aviation, retail, and telecom.
https://x.ai/news/grok-voice-think-fast-1
xAI says the model was developed in collaboration with Starlink, which is already using it for phone sales and customer support. The company shared some impressive operational metrics: 20% of callers in telemarketing campaigns purchased Starlink service on the spot; 70% of customer service requests were resolved autonomously without escalation to a human agent; and a single agent can orchestrate 28 different tools across hundreds of support and sales workflows, handling everything from hardware troubleshooting to replacements and service credit issuance.
The previous generation model, grok-voice-fast-1.0, has been marked as deprecated. xAI recommends migrating to the new model.