Anthropic's AI just crushed a record: 9 minutes vs. 181 for humans controlling a robot

The ‘bitter lesson’ is an old chestnut in AI: that methods leveraging massive computation ultimately triumph over clever, hand-crafted human approaches. This week, Anthropic provided a spectacularly tangible demonstration of that principle, not in a simulator, but with a physical, four-legged robot.

Just nine months ago, in August 2025, the company’s Claude Opus 4.1 model was completely incapable of performing a set of real-world robotics tasks on its own. It was useful as a co-pilot for human operators, helping them work roughly twice as fast. But completing the suite of tasks still took a human-AI team a grueling 181 minutes. Fast forward to May 2026, and the latest model, Opus 4.7, operating with full autonomy, blew through nearly every task in just 9 minutes and 35 seconds. That’s an almost 20x speedup over the previous human record, achieved simply by swapping in a more powerful, scaled-up model.

The approach was the epitome of letting the bitter lesson do its work. The researchers didn’t encode sophisticated robotic control theories or intricate physical world models into the system. They relied on the raw, emergent capabilities of a scaled-up general-purpose language model. The robot’s environment was a black box, and the AI figured out the necessary interactions through brute-force intelligence. The single failure—repositioning a ball it had knocked out of place—was a task that had also stumped the human teams, suggesting a genuinely tricky physical constraint rather than a simple AI shortcoming.

This isn’t just about a faster robot. It validates the hypothesis that general-purpose intelligence can directly subsume specialized physical skills without needing a fundamental architectural rethink. The implication is a bit dizzying: as foundational models like Opus continue to scale, a whole host of real-world physical tasks, from warehouse logistics to disaster response, might become solvable by the same software that writes our emails. The limiting factor is shifting from algorithmic cleverness to the simple, relentless march of compute and data.