xAI debuts Grok 4, "smartest AI in the world"
Add Axios as your preferred source to
see more of our stories on Google.

Screenshot: xAI on X
Elon Musk unveiled the newest edition of xAI's flagship AI model, Grok, late Wednesday night in a livestream video that touted Grok 4's prowess at topping benchmark scores.
Why it matters: xAI is in an accelerating race with OpenAI, Anthropic and other players to deliver faster, more reliable AI and fulfill the wild expectations investors now have for the technology.
Driving the news: Musk and several xAI researchers showed Grok 4 solving an advanced math problem, generating an image of two black holes colliding, and predicting the most likely winner of next year's World Series (the Dodgers at a 21.6% chance).
- xAI is claiming Grok as the new top-scoring model on Humanity's Last Exam, a collection of very advanced problems, as well as on other yardsticks of AI progress.
Between the lines: Grok 4 includes a new "heavy" version that deploys multiple agents to collaborate on solving particularly tough problems.
- Grok 4 is available to users immediately. The heavy mode is available at a $300 monthly rate.
The xAI researchers also demoed a new Grok 4 voice mode that's more responsive and natural than Grok's current one.
- But the voice model faltered when asked to whip up an "opera about Diet Coke," delivering what sounded more like a singsong Shakespeare monologue instead.
What they're saying: Musk repeatedly expressed amazement at the accelerating rate of Grok's advances — which, in his telling, seemed to be piling up even as he spoke.
- "Grok 4 is smarter than almost all graduate students in all disciplines simultaneously," Musk first declared.
- A few minutes later, he said, "Grok 4 is post-graduate, Ph.D. level in everything."
- A minute later, Grok had become "better than Ph.D. level."
What they're not saying: Musk and xAI's team asked Grok to solve a variety of problems in their hourlong demo.
- But they didn't once mention Grok's extraordinary meltdown on Tuesday, when the bot began spouting antisemitic ideas and calling itself "MechaHitler."
- Earlier Wednesday, Musk had addressed that separately, saying the platform had erred because it was "too eager to please" and that it would be fixed.
What's next: Musk and his team promised a new coding model in August, a multimodal agent in September and a video-generating model in October.
