In August 2025 Anthropic ran Project Fetch, pitting employees with no robotics background against a quadruped robot using Claude Opus 4.1. One team received model assistance while a control group relied on web searches and ingenuity. The Claude‑enabled group completed more tasks, faster, proving the LLM could boost non‑expert robot operation, demonstrating that language models can act as interim robot operators.

Phase Two replaced the human‑assisted team with Claude Opus 4.7, which ran tasks autonomously after a laptop connection. The model completed every measurable step at least ten times quicker, and on the four tasks finished by both original teams it averaged 20 times faster than the top human group. Code size dropped to one‑tenth of the human output, despite occasionally using an outdated detection algorithm.

Claude still fails at fine‑grained ball handling; humans nudged the beach ball reliably, while the model only managed clumsy pushes. The experiment shows LLMs can now replace pair‑programming for many robot‑interface tasks, but closed‑loop control remains a human strength. Anthropic concludes that autonomous physical agents are emerging, yet robust tool‑building and policy generation still lag behind.