xAI Releases Grok Voice 3.0, Tops Speech-to-Speech Benchmark

Beating AI News Flash: xAI announces the launch of its speech-to-speech model Grok Voice 3.0, claiming it ranks first on the Speech-to-Speech Index independently tested by Artificial Analysis with a score of 86.9, higher than Grok Voice 2.0's 81.3, GPT-Live-1's 83.2, and Gemini 3.8 Live's 82.6; it also claims end-to-end latency of under 1 second. The model is now live via the Voice Agent API, priced at $0.08 per minute, the same as its predecessor, and agents using grok-voice-latest will automatically switch over.

Officially, Grok Voice handles millions of conversations daily at Tesla and SpaceX, with the Starlink customer service center receiving approximately 20,000 calls per day, about 70% of which are resolved without human intervention, and about a quarter of sales calls completing a purchase before hang-up. It achieves 74% accuracy on ServiceNow EVA-Bench and 71.6% on τ-Voice, which is claimed to be comparable to GPT-Live-1 but with roughly 40% lower latency.