Claude Opus Breaks nanoGPT Speedrun Record: AI vs. Human

Claude Opus 4.7 recently broke a human world record on the nanoGPT speedrun, an AI training benchmark. It achieved a training configuration in 2,930 steps, surpassing the human record of 2,990 steps. OpenAI's Codex also beat the human mark, reaching 2,950 steps. These models operated within an automated pipeline, without direct human intervention. The nanoGPT speedrun challenges researchers to train a specific AI model as fast as possible. While Claude Opus 4.7 showed this impressive performance, a larger study suggests it's an isolated win. Here's the thing: this single success required extensive computing power, using 10,000 runs and 14,000 GPU-hours over two weeks. What's interesting is that humans achieved similar results with significantly less resources. A broader evaluation by METR found that AI agents, given a substantial budget, recovered less than 10% of the speedup humans achieved in the same timeframe. The bottom line is that while AI can outperform humans in specific, narrow tasks, the real gap lies in generating new algorithmic ideas, where human innovation still leads.