SpaceXSI released Grok Voice 3.0 on October 14. The speech-to-speech model is available through its Voice Agent API at the same price as its predecessor.

Grok led one broad benchmark, while OpenAI scored higher on a test built around difficult customer-support calls. For companies choosing AI for phone support, the split shows why one leaderboard cannot settle which model will work better for their customers.

Different tests, different leaders

In Artificial Analysis’s Speech-to-Speech Index, Grok Voice 3.0 scored 86.9. The comparison reported by SpaceXSI also ranked GPT-Live-1 and Gemini 3.8 Live:

  • Grok Voice 3.0: 86.9
  • GPT-Live-1: 83.2
  • Gemini 3.8 Live: 82.6

The index combines measures such as speech reasoning, task success and listener preferences in blind comparisons. It does not guarantee that a model will perform best in every business or live call.

A different result emerged on τ-Voice, which tests support calls involving accents, background noise and interruptions:

  • GPT-Live-1: 74.5%
  • Grok Voice 3.0: 71.6%

SpaceXSI says Grok Voice 3.0 reached its first audio about 40% faster in that test. The figures point to a tradeoff: one model ranked higher on the reported support-call score, while the other was faster to begin speaking.

What the Starlink figures show

SpaceXSI says Starlink uses Grok Voice to handle about 20,000 customer calls a day in multiple languages. The company says roughly 70% are resolved without a human. These are company-reported figures, not an independent audit.

The Voice Agent API costs $0.08 per minute, the same listed rate as Grok Voice 2.0. SpaceXSI’s release notes say developers using the grok-voice-latest alias have been switched to version 3.0 automatically.

For call-center buyers, the useful test is not just which model tops a benchmark. It is how well it handles their customers, call conditions and required handoffs.

The published Starlink figures offer one example of deployment at scale, but they do not guarantee the same results for another company. Buyers will need to test the model against the calls they actually handle.