xAI has released Grok Voice 3.0, a voice model that takes in spoken audio and answers in spoken audio, and the strongest evidence for it comes from outside the company. Artificial Analysis, the independent benchmarking firm, puts it first on its Speech-to-Speech Index with 86.9. GPT-Live-1 scores 83.2, Gemini 3.8 Live 82.6, and xAI’s previous version, Grok Voice 2.0, 81.3.
The jump over 2.0 is real on the index, but it is uneven underneath. Task success moved from 94.6 percent to 94.8 percent, a gap of 0.2 points. Big Bench Audio, a test of reasoning over speech, went from 97.2 to 98.5 percent. The larger margins are against rivals: 94.8 percent task success versus 87.4 for GPT-Live-1 and 89.1 for Gemini 3.8 Live. On Full Duplex Bench, which scores how well a model handles pauses and interruptions, Grok Voice 3.0 reaches 97 percent.
The customer-support claim is harder to check. xAI says the model “sets a new record for resolving customer support calls,” and points to τ-Voice, a test built around accents, background noise, and callers who interrupt. Grok Voice 3.0 scores 71.6 percent there, which xAI itself describes as comparable with GPT-Live-1. The announcement prints no τ-Voice figure for Grok Voice 2.0 or any other rival, so the record cannot be read off the page. A second result, from ServiceNow’s open-source EVA-Bench for enterprise voice agents, gives Grok Voice 3.0 the top accuracy score at 74 percent. A ServiceNow product executive, Nitin Aggarwal, is quoted praising it. The announcement does not say who ran that evaluation.
Latency has the same gap. xAI calls the model “class-leading” at under one second, but gives no measurement point (first sound heard, or full round trip), no hardware, and no count of simultaneous calls. The only comparison is a claim of 40 percent lower latency than GPT-Live-1 at similar τ-Voice accuracy, with no absolute time for either.
Price is the cleanest fact. The model is in xAI’s Voice Agent API at $0.08 a minute, or $4.80 an hour, the same as Grok Voice 2.0. Developers who point at the grok-voice-latest alias have already been moved to the new model automatically. Artificial Analysis also measured cost per task on Big Bench Audio: $0.025 for Grok Voice 3.0, $0.030 for GPT-Live-1, and $0.055 for GPT-Realtime-2.1. That is under half the cost of the most expensive rival, though it is a per-task figure and the announcement lists no competitor per-minute prices. It names no languages beyond saying Starlink’s calls span several.
The deployment numbers all come from Tesla and SpaceX, the Elon Musk companies xAI names as customers. Inside Tesla vehicles, xAI says, Grok Voice takes part in millions of daily conversations. Starlink call centers use it for 20,000 calls a day. Roughly 70 percent of those are resolved without a human, and a quarter of the sales calls result in a purchase. These are xAI’s figures, with no baseline and no definition of “resolved.”
For anyone on the alias, the model underneath changed without a decision on your side. The index is a fair first filter, but a support line with its own accents, noise, and angry callers should replay recorded calls through it before trusting a ranking built on someone else’s.
Based on xAI’s Grok Voice 3.0 announcement page, retrieved 15 October 2026 (the page shows no publication date).
