xAI's Grok 4 Sets New Benchmark Records

xAI, the artificial intelligence company founded by Elon Musk, has released Grok 4, a frontier large language model that claims top scores across several major AI benchmarks. The model achieves 93.8% on MMLU, 96.2% on HumanEval, and 97.1% on the MATH benchmark, edging past GPT-5 and Claude Opus 4.6 on these measures.

Grok 4 was trained on xAI's Colossus supercomputer in Memphis, Tennessee, which now houses over 200,000 NVIDIA H100 GPUs. The company says the expanded compute cluster allowed it to scale training significantly beyond Grok 3, with particular focus on mathematical reasoning and code generation capabilities.

Beyond benchmarks, Grok 4 introduces a "real-time knowledge" feature that integrates live data from X (formerly Twitter) and other sources directly into its responses. This gives the model an edge in current events queries, though accuracy of real-time information remains an ongoing challenge. xAI claims a 94% factual accuracy rate on current events questions.

The model is immediately available to X Premium+ subscribers ($16/month) through the Grok interface on X, and through the xAI API for developers. Enterprise pricing undercuts competitors significantly at $10 per million input tokens, roughly half the cost of comparable models from OpenAI and Anthropic.

Independent evaluators have begun testing Grok 4, with early results broadly confirming xAI's benchmark claims. However, some researchers note that benchmark performance increasingly diverges from real-world utility, and that head-to-head human preference evaluations may tell a different story. Full third-party evaluations are expected within the coming weeks.