xAI Releases "Grok Voice Think Fast 2.0" With Major Gains in Accuracy and Speed

xAI announced a new version of its voice model, “Grok Voice Think Fast 2.0.” Coming less than three months after the original “Think Fast 1.0” debuted in April 2026, it’s designed to listen, reason, and speak simultaneously, aiming for fluid conversation with no awkward pauses or waiting.

Details

• Response speed: Time to first audio response is 0.70 seconds. Production tool calls often begin executing before the agent even finishes its first sentence
• Benchmarks: Scored 82.9% overall on Artificial Analysis’s Speech-to-Speech Quality Index (up from 75.7% for v1.0) — ahead of GPT-Realtime-2.1 (79.1%) and Gemini 3.1 Flash (69.5%)
• Transcription accuracy: Claims a 1.5-2.0x improvement over Deepgram Nova 3 and ElevenLabs Scribe v2, and roughly 10x better accuracy than dedicated transcription models in noisy environments
• Language support: Supports more than 25 languages, with strengthened handling of noisy environments and varied accents
• Efficiency: Reasoning-token usage per response fell to 0.4x the previous version (roughly a 40% reduction)
• Pricing: $0.08 per minute of audio

How to try it

• Starting August 5, 2026, grok-voice-latest will automatically move from grok-voice-think-fast-1.0 to grok-voice-think-fast-2.0 — no action needed for existing implementations
• Users who want to stay on the previous version can explicitly pin grok-voice-think-fast-1.0 before the switchover