Grok Voice 3.0: xAI's voice agent goes sub-second at flat pricing

What xAI released

On October 14, 2026, xAI announced Grok Voice 3.0, its latest speech-to-speech model for live voice agents, available immediately in the Voice Agent API. Pricing holds at $0.08 per minute ($4.80 per hour), unchanged from Grok Voice 2.0, and agents pinned to grok-voice-latest were migrated automatically with no developer changes required. xAI frames the release around three claims: leading speech-to-speech intelligence, sub-second latency, and lower running cost than comparable frontier models.

According to Artificial Analysis measurements cited by xAI, the average cost per task in the Big Bench Audio run was $0.025 for Grok Voice 3.0, $0.030 for GPT-Live-1 (Astra, Medium), and $0.055 for GPT-Realtime-2.1 (High). These benchmark per-task inference costs differ from the $0.08-per-minute API price. They provide a reference; actual call-center economics depend on workload, call length, integrations and human escalation.

[1] [2]

The scoreboard, as xAI presents it

xAI reports that independent testing firm Artificial Analysis ranks Grok Voice 3.0 first on its Speech-to-Speech Index at 86.9 -- an index combining speech reasoning, tau-Voice support-call tasks, and preference plus task success from blind head-to-head calls. The published comparison table puts GPT-Live-1 (Astra, Medium) at 83.2, Gemini 3.8 Live (Extended Thinking, High) at 82.6, and Grok Voice 2.0 at 81.3. Component scores show where the gains came from: speech reasoning on Big Bench Audio rose to 98.5% from 97.2%, Speech Agent Arena preference Elo to 1070 from 1011, task success to 94.8% from 94.6%, and Full Duplex Bench -- pauses, turn-taking, interruptions -- to 97.0% from 95.1%.

The honest wrinkle is one xAI includes itself: on tau-Voice, the benchmark built around noisy, multi-accent support calls, Grok Voice 3.0 scores 71.6% against 74.5% for GPT-Live-1 (Astra, Medium) -- it does not lead that test. xAI's framing is that the result is comparable at 40% lower latency, and that on ServiceNow's open-source EVA-Bench for enterprise voice agents it achieves the highest accuracy at 74%, ranking best when accuracy and customer experience are scored together. A ServiceNow product director is quoted calling it a new Pareto frontier for the benchmark -- an endorsement xAI selected for its own announcement, which is exactly how such quotes should be read. The 74.5% rival figure is reported by SiliconANGLE from xAI's comparison chart, not independently remeasured.

[1] [2]

Already in production

The deployment figures are the part of the announcement no benchmark can substitute. Grok Voice, xAI says, powers millions of conversations a day across Tesla and SpaceX. In Tesla vehicles, drivers use Connectors for quick hands-free requests -- checking a calendar, answering an email -- and Grok Bot for longer tasks like ordering coffee or booking dinner; Grok Voice 3.0 also powers Grok Bot's voice mode. These are company-reported figures, not independently audited, but they describe a model already embedded in shipping products rather than waiting for adoption.

Starlink's call centers are the sharper case study: 20,000 customer calls a day, across multiple languages, with roughly 70% resolved with no human in the loop, and one in four sales calls ending in a purchase before the caller hangs up. A single voice agent, xAI says, verifies customers, troubleshoots hardware, issues credits, and ships replacement equipment -- the full support workflow, not a deflection script. Again, the numbers come from the vendor. But a 70% zero-touch resolution rate, if it holds under independent measurement, is the metric that decides enterprise voice-agent budgets -- more than any index score.

[1]

Why it matters

The interesting signal is that xAI published a comparison it does not fully win. Leading the composite index while trailing on tau-Voice, then answering with latency and per-task cost, reframes the market: voice agents are now judged on a trade surface of accuracy, responsiveness, and cost per task, not a single leaderboard. A model with 40% lower latency in xAI's tau-Voice comparison and lower per-task cost in the cited Big Bench Audio run may be worth testing for production, even when a rival scores higher on one benchmark.

The deeper shift is from model quality to workflow evidence. Starlink's numbers -- resolution rate, purchase conversion, a single agent owning verify-to-ship -- are operational metrics, the kind enterprises actually budget against. Benchmarks rank models; production telemetry ranks products. xAI is betting that publishing both, on the same day, collapses the distance between best model and most practical to deploy -- and that flat pricing removes the last excuse not to try it at scale.

[1]

What to watch

Treat every vendor-cited benchmark as directional until independently replicated. The checks that matter are public: whether Artificial Analysis's published leaderboard sustains the 86.9 and the per-task cost figures, whether independent teams reproduce the 74% EVA-Bench result on the open-source harness, and whether the tau-Voice gap to GPT-Live-1 narrows in later revisions. xAI deserves credit for publishing the test it does not lead; the industry should return the favor by verifying the ones it does.

For teams running voice operations, the practical test is a pilot on your own call mix, not a leaderboard. Accent distributions, background noise, and interruption patterns vary by market, and tau-Voice exists precisely because lab scores do not transfer cleanly to real calls. The economics, though, have shifted in the buyer's favor: unchanged per-minute pricing with automatic migration means the cost of finding out is a trial, not a migration project. If 70% zero-touch resolution replicates on your calls, the benchmark debate is over.