Top AI Product

Anthropic’s Frontier Red Team likes to find where AI breaks in the physical world, and its latest test put Claude in charge of a robot dog. Project Fetch Phase Two asked whether Claude Opus 4.7 could program an off-the-shelf quadruped to sense, detect objects, and navigate — the same tasks human teams tackled a year earlier.

## What Project Fetch tested

The result was lopsided. Opus 4.7, working largely on its own, completed the challenge roughly 20x faster than last year’s best human team (which had Opus 4.1 assisting), and did it with about 10x less code — 1,045 lines that worked on the first try versus 10,309 lines that needed iteration. As a demonstration of “robotics uplift” — how much an AI raises a non-expert’s ability to build real systems — it’s a striking data point.

## Where Claude still failed

The catch is in the name: the robodog never actually fetched the beach ball. Claude struggled with closed-loop precision control — the real-time perception-and-adjust motor loop humans learn by practice. That gap is the point of a red-team exercise: code generation is racing ahead, but adaptive physical control is still where autonomous AI hits a wall.