xAI shipped a new version of its real-time speech-to-speech model, Grok Voice Think Fast 2.0, aimed squarely at developers building voice agents rather than at end users of the Grok app directly. The headline change is latency: average time-to-first-audio dropped from about 1.25 seconds in the previous version to roughly 0.70 seconds, which is the difference between a voice agent that feels like it's thinking and one that feels like it's actually listening. The "Think Fast" branding refers to the model's architecture, which reasons about a query while it is already speaking rather than pausing to think and then generating audio afterward -- a design that trades the usual latency-versus-intelligence tradeoff for something closer to both at once. xAI is also claiming meaningful gains in transcription accuracy and general conversational quality, and the model now supports automatic language detection with native-sounding accents across more than 20 languages. The part most relevant to builders is API compatibility: xAI's Speech-to-Speech endpoint is designed to be compatible with OpenAI's Realtime API, so most existing client libraries and SDKs built against OpenAI's voice endpoints should work against xAI's endpoint with little more than a base-URL change to their websocket connection. Pricing is set at $0.08 per minute of speech-to-speech audio plus a small per-token charge for text input. For teams already running voice agents on an OpenAI-compatible stack, this lowers the switching cost to essentially a config change, which is likely the point -- xAI is competing on drop-in compatibility as much as on raw model quality.