Claude Opus 4.7 controls a quadruped robot 37 times faster than humans

Anthropic has presented the results of the second phase of “Project Fetch”, an experiment that pits its flagship model Claude Opus 4.7 against human teams in controlling a quadruped robot (robodog). The data is compelling: the artificial intelligence model operated the robot completely autonomously and completed the tasks 37 times faster than the team without AI assistance, and 18 times faster than the team that did have Claude. In total, Claude Opus 4.7 was roughly 20 times faster than the fastest human team, and all of it without human intervention beyond plugging in the laptop and approving the commands. In the first phase of the project, carried out in August 2025 with Claude Opus 4.1, the human teams vastly outperformed the autonomous model, which could not even connect to the robot. However, in just ten months the leap has been transformative. Opus 4.7 not only managed to establish the connection, but also wrote effective code on the first attempt in most cases — something none of the human teams achieved — and generated ten times less code than the Claude-assisted team. According to Michael Ilie and C. Troy, researchers at Anthropic, “we are seeing a pattern in which models first help humans, and then begin to run complex tasks autonomously”. This milestone marks a turning point in the convergence between artificial intelligence and robotics. Although the model still struggles with tasks requiring millimetre precision — such as moving a beach ball with the robot — the speed and efficiency demonstrated open the door to applications in cybersecurity, industrial maintenance, exploration and hazardous environments where an autonomous agent can make decisions in real time. Anthropic already anticipates that the same dynamic that has transformed cybersecurity is beginning to materialise in the physical world. — Marta, for intermitente intelligence.