Meet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis

SpaceXAI released Grok Voice 3.0, a realtime speech-to-speech model that listens, reasons, calls tools and speaks. It scores 86.9 on the Artificial Analysis Speech-to-Speech Index, first among ranked models, ahead of OpenAI's GPT-Live-1 at 83.2 and Google's Gemini 3.8 Live Extended Thinking at 82.6. Price holds at $0.08 per minute and the endpoint is unchanged, with the latest alias already upgraded.

It isn't first everywhere: on the τ-Voice agentic benchmark it scores 71.6%, behind GPT-Live-1's 74.5%, and time to first audio of 0.81 seconds is slower than its predecessor's 0.70. Weights are closed and API-only.

Voice is where AI agents meet ordinary people — call centers, phone lines, help desks — so a better composite score at the same price matters. The jump on agentic task performance from the prior version is the number we'd circle; it's the difference between a voice that sounds good and one that actually resolves your problem.

Still, the details argue for humility. Being first on a blended index while trailing on the agentic component means buyers should test against their own call flows. And latency got slightly worse, a tradeoff that callers feel.

Two things we'd flag on the risk side. Custom voices cloned from a two-minute clip are a fraud vector in a world of phone scams, and responsibility for consent sits with whoever deploys it. And an alias that silently upgraded every existing agent is convenient but risky for regulated deployments; pinning versions should be standard practice.

The competitive picture is healthy: three major labs within a few points of each other pushes prices and quality in the right direction.