Grok Voice 3.0 Tops the Speech-to-Speech Index at 86.9 as xAI Says Starlink Handles 70% of Daily Calls

xAI has launched Grok Voice 3.0, a new speech-to-speech model that it bills as the most capable voice model available, and for once the independent benchmarking backs the claim: on the Artificial Analysis Speech-to-Speech Index, the model scored 86.9 on its first entry, taking first place ahead of OpenAI's GPT-Live-1 (Astra, medium) at 83.2 and Google's Gemini 3.8 Live Extended Thinking at 82.6. Grok Voice 2.0 scored 81.3 on the same index.

The headline latency is under one second for end-to-end response, with the company citing roughly 40% lower latency than its predecessor in voice-agent scenarios. Independent measurements come with caveats — first-audio response times and task-scenario latency are measured differently across labs, and Artificial Analysis data actually shows the new model responding a touch slower than Grok Voice 2.0 on some axes while ranking third on user preference. The task-completion story is the stronger one: Grok Voice 3.0 leads the Speech Agent Arena with a 94.8% task success rate.

Pricing is unchanged at $0.08 per minute on the Voice Agent API, which works out to about $4.80 per hour of input audio — competitive with GPT-Live-1 tiers and above Google's cheapest Gemini Live configurations. Developers using the grok-voice-latest alias will be switched to the new model automatically, and the model is also callable through xAI's Voice Agent Builder without API changes, so existing integrations upgrade by default.

The deployment evidence comes from inside Musk's own empire. xAI says Grok Voice already handles millions of conversations daily across Tesla vehicles and SpaceX, and that the Starlink support center now takes roughly 20,000 multilingual calls per day with about 70% resolved without any human involvement. The company also claims about a quarter of sales calls complete a purchase before the caller hangs up. All of those are company-reported figures from its own operations, not independent audits, but they describe one of the largest real-world voice-agent deployments in existence.

The competitive frame is shifting from voice quality to task completion. A year ago, voice models competed on naturalness and response speed; the Speech-to-Speech Index now weights agentic execution — can the model actually change a plan, process a refund, or diagnose a router fault mid-call. xAI's strategy has been to prove the loop inside its own high-volume call centers before opening it to outside developers, a closed-circuit validation approach its rivals, who license to third parties first, have less control over.

For enterprise buyers, the practical calculus is now latency, task success and cost per minute against their own workload mix. Grok Voice 3.0 leads on two of the three, at a price point that undercuts human call-center labor by an order of magnitude. The open question is whether third-party deployments — with noisier audio, angrier customers and unfamiliar tooling — reproduce the 70% containment rate xAI reports from its own frontline.