Claude Opus 4.7 programs robot dog twenty times faster than top human teams

Anthropic's Claude Opus 4.7 programmed a robot dog in Project Fetch Phase Two. The model finished every task at least ten times faster than any human team from the 2025 experiment. On the four tasks both human teams completed, it was on average more than thirty‑seven times faster than the team without Claude and more than eighteen times faster than the team with Claude. Overall it ran about twenty times faster than the quickest human‑AI team from the prior year.

The model wrote roughly one thousand forty‑five lines of code. The Claude‑assisted human team wrote over ten thousand lines. Claude Opus 4.7 achieved the same or better results while using almost ten times less code. This shows a big gain in coding efficiency.

Claude Opus 4.7 quickly linked to the robot’s video and LiDAR sensors. It chose good ways to talk to the robot and wrote control software that often worked on the first try. The model built path‑tracking and object‑detection systems and tied all parts together into a working whole.

The robot still could not reliably fetch a beach ball on its own. Precise real‑time control remains hard for the model. Low‑level actuation policies were not tested in this experiment.

The speed and code savings could let people without robotics expertise build and test robots much faster. Manufacturing, logistics and hardware research could see much shorter development cycles. This work marks an early step toward physical AI that can use off‑the‑shelf tools and robots on its own.