TechFlow news: On June 19, Anthropic announced the results of Phase Two of “Project Fetch,” evaluating performance improvements of its latest model in real-world robotic manipulation tasks. Conducted in August 2025, the experiment tasked non-robotics-expert Anthropic employees with completing a series of complex tasks using off-the-shelf quadruped robots. Performance was compared between two conditions: (1) using Claude models for assistance, and (2) relying solely on human expertise and internet resources. Results showed that under fully autonomous operation by the latest model, Claude Opus 4.7, average task completion speed across all successfully executed tasks significantly surpassed that of the human team—by at least 10×.
Anthropic emphasized that this advancement stems not from robotics-specific training but from the expanded capabilities of its general-purpose large models. The company noted that AI is transitioning—from “assisting humans in using tools” toward the early stage of “directly operating physical tools”—paralleling its earlier evolution in software engineering toward agent-based programming.
