In a move that likely has robotics software engineers across the country frantically polishing their CVs, Anthropic has revealed that its latest AI model, Claude Opus 4.7, can program a physical robot nearly 38 times faster than a team of humans. According to the company’s “Project Fetch Phase Two” research, the AI autonomously breezed through a series of complex robotics tasks in a mere 9 minutes and 35 seconds. The unassisted human team, by comparison, took a gruelling 361 minutes to finish the same job.
This isn’t just a marginal improvement; it’s a total shift in the landscape. Only ten months ago, in August 2025, Anthropic conducted the first phase of this experiment. Back then, the flagship model of the day, Opus 4.1, fell at the first hurdle, failing even to connect to the quadruped “robodog.” A human team aided by Claude took 181 minutes to complete the tasks, while those working without AI assistance struggled for over six hours. Fast forward to today, and Opus 4.7 didn’t just manage the connection; it completed the entire workflow 19 times faster than the AI-augmented humans from the original trial.
The tasks were far from trivial. They involved interfacing with the robot’s camera and lidar sensors, scripting a programme to monitor its trajectory, and employing computer vision to identify a beach ball. The human researcher’s role was reduced to the bare minimum: plugging in a laptop, providing the initial prompt, and giving the nod to the AI’s proposed actions. The AI handled the heavy lifting, from scouting the correct software libraries to writing and executing the code.
Why does this matter?
The most striking takeaway from Anthropic’s report is that this massive leap in performance wasn’t down to specialised robotics training. Instead, it is an “emergent capability”—a sophisticated byproduct of general AI scaling, the same brute-force intelligence driving improvements in chatbots and image generators. This suggests that as foundation models grow more intelligent, they will naturally become more adept at interacting with and commanding the physical world.
The technical “secret sauce” is what Anthropic calls an “agentic loop.” This is a process where the model gathers context, takes an action (such as writing a block of code), and verifies the outcome before moving on to the next step. Opus 4.7 operated with “adaptive thinking at maximum effort,” a high-intensity reasoning mode that allows the model to “think” between individual actions. This interleaved reasoning allows the AI to spot a snag—such as a dropped sensor connection—and course-correct its next command instantly, rather than grinding to a halt and waiting for a human to debug the mess. While Anthropic admits the model still finds fine-motor precision tasks a bit of a challenge, the barrier to getting robots up and running has effectively been demolished. The bottleneck is no longer the hardware; it’s about who—or what—can program it fastest. Right now, the smart money is on the silicon.
