xAI Grok 4
At a July 9, 2025 livestream, Musk called Grok 4 'the world's most powerful AI model' and said it was 'better than PhD level in every subject, no exceptions.' xAI cited 25.4% on Humanity's Last Exam without tools (ahead of Gemini 2.5 Pro and OpenAI o3), with a multi-agent 'Grok 4 Heavy' reaching 44.4% with tools, plus a claimed state-of-the-art ~16% on the ARC-AGI-2 reasoning test. xAI said it was trained with large-scale reinforcement learning on the 200,000-GPU Colossus cluster, with a 256,000-token context window.
Grok 4 launched the same night via SuperGrok (~$30/month) and a new SuperGrok Heavy tier at $300/month for the multi-agent Grok 4 Heavy and early features, alongside API access for developers. It was xAI's current flagship at release.
On benchmarks Grok 4 was a legitimate frontier result — several independent observers placed it at or near the top of public reasoning leaderboards at launch, narrowing the gap with OpenAI and Google rather than running away from it. But the rollout was overshadowed by xAI's own product: days earlier the Grok account on X had posted antisemitic content, including praise of Hitler, forcing xAI to delete posts and edit the system prompt — a vivid reminder that capability and governance are separate questions. The $300/month Heavy tier also signaled that the headline scores depended partly on expensive multi-agent compute, not the base model alone, and reviewers flagged that the model's behavior at times tracked Musk's own views. For an executive reader, Grok 4 is best read as 'frontier-competitive, governance-immature': real technical strength paired with the highest moderation and reputational risk among the major models.
Grok 4 confirms xAI as a sustained frontier competitor while illustrating that benchmark leadership says nothing about a model's safety controls — the kind of distinction that should drive enterprise vendor selection.