xAI launches Grok Voice 3.0

xAI's headline claim rests on the aggregate Speech-to-Speech Index, where Grok Voice 3.0 scored 86.9, ahead of OpenAI's GPT-Live-1 Astra (Medium) at 83.2, Google's Gemini 3.8 Live at 82.6, and its own predecessor's 81.3 ^[5]. That is a real jump. But the aggregate number hides a split result: on tau-Voice, the index component that measures whether a multi-step customer problem actually gets resolved, GPT-Live-1 Astra (medium) still leads, 74.5 percent to Grok Voice 3.0's 71.6 percent ^[3][7]. That gap has fed broader skepticism about whether Grok Voice 3.0 truly leads on the benchmarks that reflect actual human preference, given how mixed the results are across the index's individual components ^[8]. In other words, 'world's leading' is true on one specific aggregate scoreboard and contested on the metric that matters most for an actual support call.

Grok Voice 3.0's per-minute price is unchanged from Grok Voice 2.0 at $0.08 per minute, or $4.80 per hour ^[2][5]- not the kind of headline price cut that usually accompanies a new leading model. The real cost argument turns out to be more nuanced than sticker price: FourWeekMBA's analysis frames xAI's actual pitch as cost-per-completed-call rather than price-per-minute, pointing to a Big Bench Audio run where Grok Voice 3.0 cost about $0.025 per task versus $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1 ^[3][4]. That is a legitimate efficiency story if task success holds up, but it is a narrower claim than 'cheaper' - and keeping the headline per-minute price flat looks designed to ease customer migration rather than to undercut competitors on sticker price.

Away from the benchmark debate, Grok Voice is already doing production work: SpaceX's Starlink call centers field roughly 20,000 customer calls per day across multiple languages, with about 70 percent resolved without a human agent and roughly one in four sales calls converting before the caller hangs up ^[1]. ServiceNow separately ran the model against its own EVA-Bench enterprise benchmark - more than 200 scenarios - and reported 74 percent accuracy, the highest score xAI cites for enterprise-style tasks, with a product leader describing it as setting a new accuracy-versus-naturalness tradeoff frontier ^[1]. That combination of a live, high-volume deployment and an independent enterprise endorsement is arguably more consequential than the benchmark index itself, since it ties the model's performance to a measurable business outcome rather than a leaderboard position.

The release lands under a changed corporate identity: following a reported SpaceX acquisition of xAI in February 2026, the company now operates as SpaceX Super Intelligence, or SpaceXSI, and it was SpaceXSI's account that announced Grok Voice 3.0 ^[3]. That context helps explain why Starlink's call centers are the flagship deployment example rather than a third-party enterprise customer - Starlink is now effectively an in-house proving ground. It also helps explain the competitive framing against OpenAI and Google specifically: xAI timed the launch to a same-week Artificial Analysis benchmark publication and built its comparisons directly around GPT-Live-1, GPT-Realtime-2.1, and Gemini 3.8 Live ^[3][5], positioning Grok Voice as the voice layer for a combined SpaceX-xAI product stack rather than a standalone API product competing purely on price.