SpaceXSI’s Grok Voice 3.0 tops Artificial Analysis’s speech-to-speech index

SpaceXSI, which makes Grok, released Grok Voice 3.0 on 14 October and announced it on X at 22:34 IST. It is a speech-to-speech model: it hears a caller’s voice and answers in speech, for uses such as customer support lines. SpaceXSI calls it the world’s most intelligent speech-to-speech model and says it replies in under one second.

In Artificial Analysis’s independent tests, Grok Voice 3.0 ranks first on the Speech-to-Speech Index with 86.9. The index blends four measures: reasoning over speech, τ-Voice, and which model people prefer and how often it finishes the task in blind head-to-head calls.

τ-Voice tests an agent on support calls with background noise, interruptions and a range of accents. GPT-Live-1 (Astra) still scores higher on it, but SpaceXSI says Grok Voice 3.0 comes close at 40% lower latency.

Artificial Analysis also puts its average cost per task below GPT-Live-1 (Astra, Medium) and GPT-Realtime-2.1. On EVA-Bench, ServiceNow’s benchmark for enterprise voice agents, SpaceXSI reports the highest accuracy score, 74%.

In use

SpaceXSI says the model handles millions of conversations a day in Tesla cars, where drivers use it hands-free, and it powers the voice mode of Grok Bot. Starlink’s call centres take 20,000 calls a day with Grok Voice: roughly 70% are resolved without a person stepping in, and one in four sales calls closes with a purchase on the call.

Price

Grok Voice 3.0 costs $0.08 a minute ($4.80 an hour) in the Voice Agent API, the same as Grok Voice 2.0, which SpaceXSI’s pricing page now lists as deprecated. Apps that call grok-voice-latest get Grok Voice 3.0 without any code change.