Grok Voice 3.0 puts xAI on top of the speech index at $0.08 a minute
On 14 October 2026, xAI launched Grok Voice 3.0, a speech-to-speech model that takes first place on Artificial Analysis’s Speech-to-Speech Index (86.9) at an unchanged $0.08 per minute. Call centers still weighing a voice agent now have a quantified benchmark to cross against their own scenarios.
14 October 2026. xAI releases Grok Voice 3.0, its new speech-to-speech voice model. 86.9. The score that puts it first on Artificial Analysis’s Speech-to-Speech Index, ahead of GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6). $0.08. The per-minute price, unchanged from the previous generation. Why it matters: conversational voice has crossed a threshold where cost, latency, and quality converge — and one player got there without raising its price.
Speech-to-speech, not STT then TTS
The technical difference in Grok Voice 3.0 comes down to one thing: unlike the classic pipelines that transcribe speech to text, send the text to an LLM, then synthesize the reply, the model listens to spoken audio and answers directly in audio. The gain is not cosmetic; it shows up in latency — xAI claims under one second — and in how naturally the model handles pauses, turn-taking, and interruptions.
That direct architecture is what lets the model handle exactly what Full Duplex Bench measures: pause handling, turn-taking, and interruptions. Grok Voice 3.0 scores 97.0% there, against 95.1% for Grok Voice 2.0. That is the kind of metric that separates a scripted answering machine from an agent that can hold a real conversation.
The architectural point has a cost consequence worth spelling out. In a STT → LLM → TTS chain, the audio is transcribed, the transcript is billed as input tokens, the reply is billed as output tokens, and the synthesized speech is billed again as audio — three metered stages, each adding latency. A direct speech-to-speech model collapses those stages into one, which is why xAI can claim sub-second latency while keeping the price flat. For a call center running 20,000 calls a day, the difference between a three-stage pipeline and a single model is not just speed — it is the difference between an agent that interrupts naturally and one that pauses long enough for callers to hang up.
Benchmarks to read with care
The numbers xAI published are flattering, and they should be read as vendor results. On Artificial Analysis’s Speech-to-Speech Index — an equal-weighted average of four components — Grok Voice 3.0 hits 86.9, ahead of GPT-Live-1 (Astra Medium, 83.2) and Gemini 3.8 Live (Extended Thinking High, 82.6). In detail: 98.5% on Big Bench Audio (speech reasoning), an Elo of 1070 on the preference arena, and 94.8% task success.
But the index hides a nuance. On tau-Voice, a benchmark built from simulated customer-support calls, Grok Voice 3.0 scores 71.6% — behind GPT-Live-1 (74.5%), even though xAI notes its model is comparable there with 40% less latency. In other words: the xAI model leads on pure speech synthesis and audio reasoning, but OpenAI stays ahead on one specific support-task metric. The choice depends on the scenario.
Per-task cost tilts toward xAI: $0.025 on average on Big Bench Audio, against $0.030 for GPT-Live-1 and $0.055 for GPT-Realtime-2.1. On EVA-Bench, ServiceNow’s benchmark for enterprise voice agents, Grok Voice 3.0 claims the top accuracy score (74%) and “a new Pareto frontier” between accuracy and experience, in the words of Nitin Aggarwal, ServiceNow’s senior director of product management.
The deeper lesson is the gap between a benchmark and a live call. Artificial Analysis and EVA-Bench measure controlled scenarios with clean audio and a single ground-truth end state; a real caller mumbles, changes their mind, or disputes a charge. That is why xAI’s strongest evidence is not the 86.9 index score but the 70% no-human resolution rate at Starlink — a number measured on real calls, not simulated ones. Any team evaluating a voice agent should insist on the same: pilot on live traffic, not on a leaderboard.
Proof by deployment
The most credible claim is not in the benchmarks but in the production volumes. xAI says Grok Voice powers millions of conversations a day in Tesla vehicles — through Connectors for actions like checking a calendar or replying to an email, and Grok Bot for heavier tasks like ordering coffee. The model also powers Grok Bot’s voice mode.
The Starlink call centers offer a more usable number: 20,000 calls a day, handled across languages, about 70% resolved without a human in the loop, and one in four sales calls closed before the caller hangs up. A single agent built on Grok Voice 3.0 verifies the customer, troubleshoots hardware, issues credits, and ships replacement equipment. That is not a demo; it is an operational flow at scale.
The price, and what it says about the market
Grok Voice 3.0 is offered at $0.08 per minute, or $4.80 per hour — the same as Grok Voice 2.0. Agents using the grok-voice-latest alias were switched over automatically, with no code changes. Holding the price steady while capability improves is the real signal: conversational voice is becoming a commodity, and the competition now plays out on quality and latency more than on price.
For a decision-maker, the math is easy to set up. A human support agent costs, in round numbers, tens of dollars per hour loaded; a voice agent at $4.80 an hour changes the equation for simple, repetitive tickets. The question is no longer “does it work” but “which tickets can it handle alone, and where do you keep a human in the loop.”
Voice becomes a commodity, and open source lags behind
Holding the price at $0.08 a minute only makes sense in market context. OpenAI launched GPT-Live-1 at $0.05 a minute — cheaper than Grok Voice 3.0 — while Google pushes Gemini Live through its Gemini ecosystem at comparable rates. The fight is no longer about per-minute cost — all three sit in the same range — but about perceived quality, latency, and language coverage.
For a self-hoster, the picture is less cheerful. Unlike text LLMs, where open models such as DeepSeek V4.1 Flash or Qwen let you run everything locally, speech-to-speech voice remains largely closed: Grok Voice 3.0, GPT-Live-1, and Gemini Live cannot be downloaded. Mature open building blocks exist — STT and TTS models like Whisper and Piper — but assembling a low-latency speech-to-speech pipeline is still tinkering rather than a product. The gap between cloud voice and self-hosted voice has not closed; this generation widened it.
That is a data point for any sovereignty-versus-cost decision: if your constraint is conversation confidentiality or data residency, self-hosted voice still imposes an engineering cost the cloud market simply bills you by the minute. The trade-off is not neutral for a call center handling health or financial data.
Verdict
If you run a call center or a high-volume support team, test Grok Voice 3.0 on your simple-ticket scenarios — booking, identity verification, first-level diagnostics — and measure the no-human resolution rate on your data, not xAI’s. If you need leading speech reasoning or a finely tuned multilingual voice, run it against GPT-Live-1 and Gemini Live on an internal test bench. Either way, keep a human in the loop for escalations: the 70% no-human resolution rate at Starlink still leaves 30% of conversations where supervision matters.