Anthropic Project Fetch Phase Two: Claude Finishes Robot Programming Tasks Up to 37× Faster Than Human Teams
Summary: Anthropic's Frontier Red Team pitted Claude Opus 4.7 against human participants on a live robot programming challenge. The AI finished every task humans had ever completed at least 10× faster, and averaged 37× faster on the four tasks both sides attempted.
Key Facts
- Task design: Teams connected their own computers to a robot dog, tapped onboard video and lidar sensors, wrote control software from scratch, and attempted to retrieve a ball
- Claude Opus 4.7 (unassisted): ≥10× faster than any human team on all shared tasks; 37× faster on average across the four common tasks
- Most of the code Claude wrote ran correctly on the first attempt — the iterative debugging loops that slowed human teams were largely absent
- Findings draw on 400,000 interactive sessions from 235,000+ participants across the study period
- Hard limit found: closed-loop precision control using real-time visual feedback still failed — the robot dog never actually retrieved the ball
Why It Matters
Project Fetch Phase Two is one of the more controlled head-to-head comparisons of AI versus human performance on a physically grounded, open-ended engineering task. The result confirms that planning and code-generation speed are already beyond human pace — while simultaneously locating the boundary: real-time perception-action loops remain a genuine gap. That boundary matters more than the headline speed number.
Read More
- Project Fetch: Phase two — Anthropic
- Claude 20x faster on robotics tasks — FourWeekMBA
- Claude beats human teams on robotics planning — Benzinga