xAI Launches Grok Voice 3.0 With 0.81-Second Latency

xAI has released Grok Voice 3.0, a native speech-to-speech model that took first place on the Artificial Analysis Speech-to-Speech Index with a score of 86.9. The update delivers a time to first audio of 0.81 seconds, roughly 40% faster than the 1.34 to 1.35 seconds recorded by leading OpenAI and Google configurations like GPT-Live-1 and Gemini 3.8 Live, though slightly slower than Grok Voice 2.0's 0.70 seconds. xAI maintained pricing at $0.08 per minute ($4.80 per hour) across its Realtime and Voice Agent APIs. Artificial Analysis reported an average Big Bench Audio task cost of $0.025 for Grok Voice 3.0, undercutting GPT-Live-1 with Astra at $0.030 at medium settings and GPT-Realtime-2.1 at $0.055 at high settings.

On component benchmarks, Grok Voice 3.0 achieved 98.5% on Big Bench Audio reasoning, 97.0% on Full Duplex Bench, and a 94.8% task success rate in Speech Agent Arena, topping Gemini 3.8 Live at 93.2% and Grok Voice 2.0 at 94.6%. It raised its tau-Voice agentic score from 56.5% to 71.6%, trailing only GPT-Live-1 with Astra at 74.5%. However, Grok Voice 3.0 ranked third in blind human preference on Speech Agent Arena with 1070 Elo, behind Gemini 3.1 Flash Live Minimal at 1100 Elo and Gemini 3.8 Live at 1088 Elo. Overall, the model improves on Grok Voice 2.0's 81.3 index score by 5.6 points and adds 13.2 points to Grok Voice 1.0's 73.7 score from six months ago.

For practitioners, Grok Voice 3.0 maintains WebSocket session compatibility with OpenAI's Realtime API, allowing teams to migrate by switching endpoints, API keys, models, and voices. xAI automatically points the grok-voice-latest alias to version 3.0. Although input audio costs $4.80 per hour compared to $3.50 for Gemini 3.8, the service offers tool calling, mid-conversation language switching, and dozens of supported languages for enterprise phone workflows. Benchmark evaluations were run on the Rex voice out of five available personalities. Developers can connect using Python's websockets package with an XAI_API_KEY environment variable directed to wss://api.x.ai/v1/realtime?model=grok-voice-latest.