SpaceXSI has released Grok Voice 3.0, a speech-to-speech model aimed at customer support and other voice-agent jobs. The model is available through the Voice Agent API for $0.08 per minute, or $4.80 per hour, the same rate as its predecessor.
Developers already using grok-voice-latest were switched to the new version automatically, according to the company. SpaceXSI also announced access through its Voice Agent Builder alongside the API.
Faster responses, with benchmark caveats
SpaceXSI says Grok Voice 3.0 begins responding in under a second. Its announcement cites an 86.9 score on Artificial Analysis’s Speech-to-Speech Index, ahead of the compared GPT-Live-1 and Gemini 3.8 Live configurations, which scored 83.2 and 82.6, respectively.
The company also reports a 97% score on Full Duplex Bench, a test of how models handle pauses, turn-taking and interruptions. On the τ-Voice support-call test, SpaceXSI says Grok Voice 3.0 reached a 71.6% success rate and delivered 40% lower latency than GPT-Live-1 while producing a comparable result.
Those are benchmark results presented in SpaceXSI’s launch announcement, not a Digg hands-on test. The company’s broader description of the model as the “world’s most intelligent” is a marketing claim rather than a conclusion established by the launch materials.
