xAI Releases Grok Voice 3.0 with Sub-Second Latency and Industry-Leading Accuracy

xAI has released Grok Voice 3.0, a speech-to-speech model designed to set new standards for accuracy, latency, and cost in customer support and conversational AI. The model is available immediately via the Voice Agent API at the same price as its predecessor, $0.08 per minute.

Benchmark Performance

In independent testing by Artificial Analysis, Grok Voice 3.0 ranked #1 on the Speech-to-Speech Index with a score of 86.9, surpassing competitors like GPT-Live-1 and Gemini 3.8 Live. Key performance metrics include:

• Full Duplex Bench: 97.0% accuracy in handling pauses, turn-taking, and interruptions.
• Big Bench Audio: 98.5% speech reasoning accuracy.
• Task Success Rate: 94.8%.
• Latency: Class-leading time to first audio under one second.

On EVA-Bench, ServiceNow’s enterprise voice agent benchmark, Grok Voice 3.0 achieved the highest accuracy score of 74% and performed best when jointly scored for accuracy and customer experience.

Real-World Deployment

Grok Voice 3.0 is already powering millions of daily conversations across Tesla and SpaceX:

• Tesla Vehicles: Drivers use the model for hands-free tasks such as checking calendars, answering emails, and ordering coffee via Grok Bot.
• Starlink Call Centers: The model handles 20,000 customer calls daily across multiple languages. Approximately 70% of these calls are resolved without human intervention, and one in four sales calls results in a purchase.

Pricing and Availability

The model is available today in the Voice Agent API. Existing agents using grok-voice-latest have been automatically switched to the new version with no configuration changes required. The pricing remains unchanged at $0.08/min ($4.80/hr), making it cheaper per task than comparable models like GPT-Live-1 and GPT-Realtime-2.1 according to Artificial Analysis data.