Islamabad: xAI’s Grok 4, launched on July 10, 2025, has set new standards in AI performance, achieving a 44.4% score on the Humanity’s Last Exam (HLE) and outpacing rivals like Google’s Gemini 2.5 Pro and OpenAI’s o3, amid debates over its capabilities and recent controversies.
Developed by Elon Musk’s xAI, Grok 4 was unveiled as the “world’s smartest AI,” excelling in benchmarks like HLE, a 2,500-question test spanning math, sciences, and humanities, designed by nearly 1,000 experts. Grok 4 scored 25.4% without tools, 38.6% with tools, and 44.4% in its multi-agent Grok 4 Heavy mode, surpassing Gemini 2.5 Pro (26.9%) and OpenAI’s o3 (24.9%). It also achieved 87.5–88.9% on GPQA, 95% on AIME 2025, and 72–75% on SWE-Bench for Grok 4 Code, securing top spots on the Artificial Analysis Intelligence Index and ARC-AGI-2 (16.2%). Independent verification by the ARC Prize Foundation confirmed these results, though HLE leaderboard updates are pending.
The model, powered by xAI’s Colossus supercomputer with 200,000 GPUs, emphasizes reasoning through a hybrid transformer architecture and extensive reinforcement learning. However, its 128,000-token context window lags behind competitors, limiting performance on large codebases or documents like 170-page PDFs. Multimodal capabilities, including text and image processing, are also underdeveloped, with improvements planned for September 2025. A specialized coding model is slated for August, followed by video generation in October. Access costs $30/month for Grok 4 and $300/month for SuperGrok Heavy, positioning xAI as the priciest major AI provider.
Controversies, including Grok 3’s antisemitic outputs, have raised concerns about content moderation, especially as Grok 4 reportedly references Musk’s X posts on sensitive topics like immigration or the Israeli-Palestinian conflict. Despite these challenges, experts like Alex Olteanu from DataCamp praise Grok 4’s logical reasoning in math and programming, though its practical limitations and premium pricing may hinder enterprise adoption.
