Grok Voice Think Fast 2.0

Grok Voice Think Fast 2.0 became available in the xAI API in late July 2026, and from 5 August 2026 the `grok-voice-latest` alias routes to it, moving every developer off version 1.0 unless they deliberately pin the older model. The architectural point is what separates it from most voice assistants. The common approach chains three systems together — transcription, then a language model, then speech synthesis — and every link adds delay and throws away information, tone and hesitation among it. Grok Voice handles audio in and audio out within a single model, which is why xAI can quote a time to first audio of about 0.70 seconds and a score of 82.9 on the Artificial Analysis speech-to-speech quality index. The company also reports roughly 60% fewer reasoning tokens than version 1.0 for the same work — the practical meaning of the "think fast" label is that the model spends less deliberation on getting to the same answer. It is built for voice agents rather than dictation: sub-second turn-taking, tool calls including web search from inside a live conversation, and a WebSocket streaming interface alongside plain HTTP endpoints. There is a default voice named "eve", a set of built-in alternatives, and custom voices cloned from a reference clip of up to 120 seconds. xAI's wider speech stack around it covers transcription in 25 languages across 12 audio formats and text-to-speech in English, output ranging from high-fidelity MP3 down to telephony-grade μ-law. Pricing is charged by time rather than by tokens: $0.08 per minute of audio, which is $4.80 an hour, plus $0.004 for text input. That is a 60% increase over the $0.05 per minute charged for version 1.0 — the older model remains available for anyone who prefers the cheaper rate. The platform carries SOC 2 Type II, HIPAA eligibility and GDPR compliance with multi-region infrastructure.

▸Videos

No videos yet.

▸Reviews