SpaceXSI Launches Grok Voice 3.0, Tops Artificial Analysis Speech-to-Speech Index

SpaceXSI (formerly xAI) on October 14, 2026 announced Grok Voice 3.0, its latest native speech-to-speech model for the Voice Agent API. In the company’s primary post, SpaceXSI said the model targets customer-support and sales voice agents, citing class-leading latency under one second and pricing unchanged from Grok Voice 2.0.

Independent evaluator Artificial Analysis ranked Grok Voice 3.0 #1 on its Speech-to-Speech Index at 86.9 (as of October 13, 2026), ahead of GPT-Live-1 (Astra, Medium) at 83.2 and Gemini 3.8 Live (Extended Thinking, High) at 82.6. Per SpaceXSI’s published table, the model scored 98.5% on Big Bench Audio speech reasoning, Elo 1070 on Speech Agent Arena human preference, 94.8% task success rate, and 97.0% on Full Duplex Bench. On τ-Voice, it scored 71.6%—behind GPT-Live-1 Astra’s 74.5% in Artificial Analysis’s reporting—while SpaceXSI said it reached first audio about 40% faster on that test.

On ServiceNow’s EVA-Bench for enterprise voice agents, SpaceXSI said Grok Voice 3.0 posted the highest accuracy at 74% and led when accuracy and customer experience were scored together. ServiceNow’s Nitin Aggarwal was quoted saying the model sets a new Pareto frontier on that benchmark. SpaceXSI also said Grok Voice powers millions of Tesla in-car conversations daily and answers about 20,000 Starlink call-center calls per day across languages, with roughly 70% resolved without a human and one in four sales calls ending in a purchase—company operational claims DigiEditorial has not independently audited.

Grok Voice 3.0 is available now in the Voice Agent API at $0.08 per minute ($4.80 per hour), the same price as Grok Voice 2.0. Agents already pointed at the grok-voice-latest alias were switched automatically. DigiEditorial verified the launch against SpaceXSI’s October 14 announcement at x.ai/news/grok-voice-3, with corroboration from SiliconANGLE and heise online. Benchmark rankings and EVA-Bench figures are as reported by Artificial Analysis and SpaceXSI; DigiEditorial has not independently re-run those evaluations.