Evergreen Tools
|
← Back to AI Tools

Grok Voice 3.0: A Sub-Second Speech-to-Speech Model

xAI released Grok Voice 3.0 on 14 October 2026, a speech-to-speech model for the Voice Agent API that it says answers in under one second. On Artificial Analysis's Speech-to-Speech Index it scores 86.9 for first place, 97% on Full Duplex Bench, and 71.6% on tau-Voice at 40% lower latency than GPT-Live-1, priced at $0.08 per minute like its predecessor with grok-voice-latest already pointing at 3.0

Tool Interface

Interactive tool will be available soon

Features

  • ✓ A single speech-to-speech model with sub-second latency aimed at real-time support and voice assistants
  • ✓ First place on the Artificial Analysis Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 at 83.2 and Gemini 3.8 Live at 82.6
  • ✓ 97% on Full Duplex Bench, which measures how naturally it handles pauses, turn-taking and interruptions
  • ✓ Highest accuracy on ServiceNow's EVA-Bench at 74%, and xAI says it also leads when accuracy and customer experience are scored together
  • ✓ $0.08 per minute ($4.80 per hour), unchanged from Grok Voice 2.0, and agents on grok-voice-latest were switched automatically

How to Use

  1. Create a Voice Agent in the xAI console and pick grok-voice-latest, which already points at 3.0
  2. Test with accented, noisy and interruption-heavy samples rather than a quiet studio recording
  3. Budget at $0.08 per minute and keep text input plus server-side search tools as separate line items
  4. Update any code that hard-codes grok-voice-2.0, since that model id is marked deprecated

FAQ

When was Grok Voice 3.0 released?

14 October 2026, in xAI's post 'Introducing Grok Voice 3.0'. Source: https://x.ai/news/grok-voice-3

Who measured the 86.9 score?

Artificial Analysis ran the independent testing, and xAI cites results dated 13 October 2026. It is third-party data quoted by the vendor, useful for direction but not an absolute verdict.

Did pricing change?

No. It is $0.08 per minute ($4.80 per hour), the same as Grok Voice 2.0, with text input billed at $0.004 and server-side search tools charged on top.

Does it hold up in production?

xAI cites Starlink call centres handling 20,000 calls a day with roughly 70% resolved without a human, and millions of conversations a day in Tesla vehicles. Those are vendor numbers and case studies, not third-party audits.

Do I need to change code when upgrading from 2.0?

Agents using grok-voice-latest were switched over already with no changes. Code that hard-codes grok-voice-2.0 needs updating because that model id is deprecated.