xAI released Grok Voice Think Fast 2.0, a speech-to-speech model built for developers building voice agents, priced at $0.08 per minute of audio. The company says the update raises performance across most use cases without requiring any prompt changes on the developer side. That claim matters because Grok Voice already handles customer-facing phone calls, and a quiet quality shift changes what those systems say to callers.
Think Fast 2.0’s overall score on Artificial Analysis’ speech-to-speech benchmark reached 82.9%, a jump from 75.7% for the first Think Fast release. That’s ahead of GPT-Realtime-2.1’s 79.1% and Gemini 3.1 Flash’s 69.5%. The agentic subscore rose to 56.5%: the predecessor managed 52.1%, GPT-Realtime-2.1 reached 45.7%, and Gemini 3.1 Flash trailed at 37.7%. First-audio latency, how long a caller waits before the model starts speaking, dropped to 0.70 seconds from 1.25 seconds. One benchmark went the other way: Think Fast 2.0’s conversational score of 95.1% sits just under GPT-Realtime-2.1’s 95.7%.
Word error rate, the standard measure for transcription quality where lower is better, improved by 1.4 times over the first Think Fast and by 1.5 to 2.0 times over two competing speech-to-text tools, Deepgram’s Nova 3 and ElevenLabs’ Scribe v2, according to xAI’s internal test across 24 languages. In noisy conditions, phone calls with compression and background noise, the company puts the gap at roughly ten times.
Instead of pausing to think, Think Fast 2.0 processes reasoning alongside speech generation, which xAI says preserves latency on harder queries. Median reasoning-token use dropped to 0.4 times the predecessor’s baseline, and xAI says that gives most tool calls time to fire before the model has finished speaking its first sentence. A reinforcement-learning pass also nudged the model toward brief responses, one question per turn, with less filler while it walks a caller through a multi-step task.
xAI points to an internal test on Starlink’s phone line: callers converted to sales more often and stayed contained in automated support more often, the company says, without publishing the underlying figures. That leaves the improvement unverified outside xAI’s own account, which TestingCatalog covered on July 29.
The detail that matters for anyone running Grok Voice in production: the grok-voice-latest alias switches automatically from Think Fast 1.0 to Think Fast 2.0 on August 5, 2026. Any integration pointed at that alias, rather than a pinned model identifier, inherits the new model’s behavior on that date with no deploy, no code change, and no warning. Teams that rely on tuned prompts, latency assumptions, or call scripts validated against 1.0 should pin grok-voice-think-fast-1.0 explicitly before August 5, then test 2.0 against their own call flows on their own schedule. Everyone else needs to do nothing, which is exactly the setup that produces surprise regressions when nobody remembers who owns the integration.
TestingCatalog reported the Grok Voice Think Fast 2.0 launch on July 29, 2026.
