xAI flips Grok Voice's default to a reasoning speech-to-speech model
grok-voice-latest now routes to Think Fast 2.0 at $0.08/min: 0.70s to first word, sharper transcription, 24 languages. Pin 1.0 to keep the old behavior.
Copy markdown
The default flips today — pin 1.0 or ride 2.0
As of August 5, `grok-voice-latest` stops resolving to Grok Voice Think Fast 1.0 and routes every call to Think Fast 2.0 automatically. If your voice agent leans on 1.0's timing or behavior, pin `grok-voice-think-fast-1.0` explicitly before your next deploy — otherwise you inherit the new model with no code change.
One model that reasons while it talks
Think Fast 2.0 is native speech-to-speech: it reasons through your query in parallel with speaking instead of pausing to think first, so the reasoning adds no latency. It's live now at $0.08 per minute of audio via console.x.ai.
Faster first word, sharper ears
Time to first audio drops to 0.70s from 1.25s. Transcription is 1.5–2.0x better than Deepgram Nova 3 and ElevenLabs Scribe v2 (1.4x over v1.0), with roughly 10x gains on noisy audio, measured across 24 languages.
What you can build this weekend
Sub-second, reason-while-speaking voice makes real-time agents — phone support, live tutors, hands-free copilots — feel like conversation, not walkie-talkie. Collapse a Whisper→LLM→TTS chain into one $0.08/min endpoint and delete two hops of latency.
Elsewhere: DeepSeek V4 Flash is 90% off on Vercel's gateway
Through Novita on Vercel's AI Gateway, the open MIT-licensed DeepSeek V4 Flash is 90% cheaper for Pro users through August 11 — a near-free window to wire it into an agent loop. Vercel also shipped project-level spend caps with email alerts.