Summary: On October 14, 2026, xAI announced Grok Voice 3.0, a speech-to-speech model it positions as the most accurate on the market, with sub-one-second latency and pricing unchanged from its predecessor at $0.08 per minute. The release matters less as a leaderboard win than as a drop-in upgrade: agents pinned to grok-voice-latest were switched over automatically. Here’s what the numbers say and what they mean if you’re building voice agents.
The Numbers, Attributed
The benchmark claims come from third parties, which is worth stating up front. Per xAI’s announcement, Grok Voice 3.0 ranks #1 on Artificial Analysis’s Speech-to-Speech Index with a score of 86.9, ahead of Grok Voice 2.0 (81.3), GPT-Live-1 Astra Medium (83.2), and Gemini 3.8 Live (82.6). It scores 97% on Full Duplex Bench, which measures how naturally a model handles pauses, turn-taking, and interruptions — the stuff that makes callers feel like they’re talking to someone rather than waiting for a turn.
On ServiceNow’s EVA-Bench, an open-source benchmark covering 200+ enterprise support scenarios, the model scored 74% accuracy. Nitin Aggarwal, Senior Director of Product Management at ServiceNow, described it as setting “a new Pareto frontier” on accuracy and experience jointly.
The claim I find most useful for planning: xAI says the model hits τ-Voice scores comparable to GPT-Live-1 (Astra) at 40% lower latency. That’s the classic voice-agent tradeoff — reasoning quality versus responsiveness — and if it holds in your own calls, it changes what you can attempt in a single turn.
The Production Evidence Is the Interesting Part
Benchmarks get you a shortlist. What should move the model up your evaluation list is where it already runs. xAI says Grok Voice powers millions of daily conversations across Tesla and SpaceX. In Tesla vehicles, drivers use it hands-free for calendar checks, email, and bigger tasks like ordering food; the announcement notes Grok Voice 3.0 also powers Grok Bot’s voice mode.
The Starlink call-center deployment is the more instructive case for builders. Per xAI, Starlink call centers answer 20,000 calls a day across multiple languages, with roughly 70% resolved without a human, and one in four sales calls ending in a purchase before hang-up. A single agent verifies customers, troubleshoots hardware, issues credits, and ships replacements. That’s an end-to-end agentic workload, not a demo — the same pattern of agents picking up real operational work I looked at in five launches that show agents doing real work.
What the Unchanged Price Actually Signals
Grok Voice 3.0 is available today in the Voice Agent API at $0.08/min ($4.80/hr), identical to Grok Voice 2.0, and agents using grok-voice-latest were migrated automatically with no code changes. Artificial Analysis’s per-task cost measurement on Big Bench Audio puts it below both GPT-Live-1 (Astra, Medium) and GPT-Realtime-2.1, per xAI.
Holding price flat while improving accuracy and cutting latency is a deliberate move: it makes the upgrade decision trivial for existing customers and pressures competitors on cost-per-resolved-call rather than cost-per-minute. For your roadmap, the practical takeaway is to budget voice agents in cost per completed task, not per minute — a model that resolves more calls in fewer turns wins even at the same rate card.
A Grounded Next Step
Vendor benchmarks and vendor-cited deployments aren’t a substitute for your own evals. If you already run a voice agent, the cheapest experiment available is repointing it at grok-voice-latest and comparing resolution rates and interruption handling on real transcripts. If you’re starting fresh, EVA-Bench is open source and worth running against your own call scenarios before you commit. The one caveat: the supplied announcement doesn’t publish regional availability limits or rate-limit details, so check the docs before promising latency numbers to a customer.