Project Fetch: Claude Opus 4.7 commands a robot without human assistance

Anthropic's Claude Opus 4.7 robotically executes autonomous tasks 20 times faster than humans in the second phase of Project Fetch, using Claude Code.

Anthropic has released the second phase of Project Fetch, an experiment by its Frontier Red Team on the use of AI-assisted off-the-shelf robots. In the first phase, last year, employees without robotics expertise, randomly assigned to groups with or without Claude, had to pilot a robotic quadruped, a robodog. The team assisted by Claude (then Opus 4.1) outperformed the other, but the model was unable to complete the tasks autonomously, already struggling with connecting to the robot.

This time, Claude Opus 4.7, operating autonomously within Claude Code, executes most of the tasks without assistance. Anthropic reports that it was approximately twenty times faster than the best human team from the previous experiment, and, on the four tasks common to both teams, more than eighteen times faster than the team already assisted by Claude, while producing nearly ten times less code. The human role was limited to plugging in the laptop, providing the initial prompt, and approving the commands.

The company emphasizes the limitations: the models have not solved robotics by any means. Claude still struggles with precise ball movement, that closed loop of perception and correction where humans remain superior, and no task involves low-level control, such as writing an actuation policy. Anthropic also notes that this progress is not the result of a dedicated effort in robotics, but emerges from general scaling. Beyond the robot, it sees a pattern already observed in cybersecurity, with models first assisting humans, then humans assisting models, before models largely act autonomously, and believes that an emerging era of physical agentic AI is taking shape.