xAI launches Grok Voice 3.0, keeps price at $0.08 a minute

Summary

• xAI released Grok Voice 3.0 on October 14, 2026, through its Voice Agent API.
• Pricing stays at $0.08 per minute, identical to Grok Voice 2.0's rate.
• Starlink call centers resolve about 70% of 20,000 daily calls without humans.

The latest

xAI put its new speech-to-speech model into commercial hands on October 14, 2026, pricing Grok Voice 3.0 at $0.08 per minute — the same rate its predecessor carried. The company calls it the world's most intelligent speech-to-speech model and reports sub-second response times. The grok-voice-latest alias in the API switches to it automatically.

Details

• The benchmark: xAI published Artificial Analysis measurements dated October 13, 2026, placing Grok Voice 3.0 first on the AA Speech-to-Speech Index at 86.9, against 83.2 for GPT-Live-1 (Astra, Medium), 82.6 for Gemini 3.8 Live (Ext. Thinking, High), and 81.3 for its own Grok Voice 2.0.
• Reasoning scores: On Big Bench Audio, which tests audio reasoning, the model scored 98.5%, ahead of Gemini 3.8 Live at 97.7%, Grok Voice 2.0 at 97.2% and GPT-Live-1 at 90.1%. Task Success Rate came in at 94.8% and Full Duplex Bench at 97.0%, both the highest in the published set.
• Where it trails: The same table shows Grok Voice 3.0 behind on τ-Voice, the agentic success measure, scoring 71.6% against 74.5% for GPT-Live-1 (Astra, Medium). xAI frames the gap through latency instead, saying its τ-Voice response time runs 40% below GPT-Live-1, without publishing timing figures beyond sub-second.
• Human preference: On Speech Agent Arena Elo, which ranks models by human preference, Grok Voice 3.0 scored 1070 — above GPT-Live-1 at 1048, Grok Voice 2.0 at 1011 and Gemini 3.8 Live at 990. It is the only metric in the set built on listener judgement rather than automated scoring.
• Cost per task: xAI puts the cost of a Big Bench Audio task at $0.025 for Grok Voice 3.0, against $0.030 for GPT-Live-1 (Astra, Medium) and $0.055 for GPT-Realtime-2.1 (High) — roughly 55% less than the most expensive comparison in the table.
• Enterprise testing: On ServiceNow's EVA-Bench, which runs more than 200 enterprise scenarios, the model posted the highest accuracy score at 74% and the best combined accuracy and customer-experience result. Nitin Agarwal of ServiceNow said "Grok Voice 3.0 sets a new Pareto frontier on EVA-Bench".
• Starlink deployment: Starlink call centers handle 20,000 customer calls a day across multiple languages, with roughly 70% resolved without a human agent, xAI said. One in four sales calls ends in a purchase. The company did not break out which languages are covered or how the resolution rate is measured.
• Tesla deployment: Tesla runs millions of in-car conversations daily through hands-free Connectors and Grok Bot tasks, according to the announcement. The same model powers voice mode inside Grok Bot, making the two deployments the largest usage figures xAI disclosed.
• What's unspecified: xAI said the model is available now in the Voice Agent API but did not name supported regions or publish a language list. Arabic is not mentioned anywhere in the announcement, and the multilingual claim appears only in reference to Starlink's call operations.

Between the lines

Holding the price at $0.08 a minute while claiming higher benchmark scores shifts the pitch from capability to unit economics. The Starlink figure — 70% of 20,000 daily calls closed without staff — functions as a live reference deployment rather than a lab result, which is the harder number for enterprise buyers to dismiss. The τ-Voice shortfall against GPT-Live-1 sits in the same table xAI chose to publish.

What's next

Watch whether Artificial Analysis or independent testers reproduce the 86.9 index score, whether xAI publishes supported regions and languages, and whether rival providers cut per-minute pricing in response.

Source: xAI