Anthropic Project Fetch Phase Two

Claude Opus 4.7 Outperforms Human Teams in Robotic Control

Anthropic's Project Fetch Phase Two reveals that Claude Opus 4.7 can autonomously complete robotic programming and setup tasks significantly faster than human teams. In comparative tests using an off-the-shelf robotic quadruped ("robodog"), the model was approximately 20 times faster than the fastest human team across all completed tasks, demonstrating a transition from AI acting as a human assistant to AI operating independently in physical environments.

Experimental Design and Methodology

Project Fetch Phase Two evaluated the autonomous capabilities of Claude Opus 4.7 using Claude Code with adaptive thinking set to maximum effort. The experiment focused on a subset of tasks previously performed by humans in Phase One, including:

• Connecting to the robodog's video camera.
• Connecting to the robodog's lidar sensor.
• Developing a program to manually control the robodog.
• Writing a program to detect a beach ball.
• Monitoring the robodog's path through space.

Researchers limited their role to providing the initial prompt, plugging in the laptop, and approving commands and task transitions. The model was not evaluated on its ability to use a physical controller, nor was the time to retrieve a ball using a Claude-programmed controller measured.

Key Performance Benchmarks

Claude Opus 4.7 showed substantial improvements in speed, efficiency, and reliability over the human teams from the previous phase:

Speed and Efficiency

For the four tasks completed by all participants, Opus 4.7 completed them in 9 minutes and 35 seconds. In comparison, "Team Claude" (humans assisted by Claude Opus 4.1) took 181 minutes, and "Team Claude-less" (humans without AI assistance) took 361 minutes. This makes Opus 4.7 approximately 18.9 times faster than the AI-assisted human team and 37.7 times faster than the unassisted human team.

Code Volume

Opus 4.7 produced significantly less code to achieve its goals, writing 1,045 lines of code compared to 10,309 lines written by Team Claude. This suggests the model was more effective at identifying the optimal path to interface with sensors and writing functional code on the first attempt.

Reliability

Performance times remained relatively consistent across three trials, though some variance occurred due to the model occasionally selecting suboptimal or outdated object detection algorithms.

Current Limitations in Physical Control

Despite its success in programmatic setup, Claude Opus 4.7 struggled with high-precision physical interaction, specifically the "fetching" part of the project. The model was unable to precisely nudge a beach ball back to a starting base, a task requiring a closed-loop feedback system to perceive errors and adjust inputs in real-time.

Anthropic notes that while humans can learn this skill with practice, the model's efforts were poorly controlled and unsuccessful. However, the lab observed that a researcher with robotics experience could successfully program autonomous fetching, suggesting that with more scaffolding, current Claude models could eventually achieve this.

Implications for Physical Agentic AI

Anthropic concludes that these results are an emergent property of general scaling rather than a specific effort to improve robotics capabilities. The findings suggest a trajectory toward "physical agentic AI," where models can use off-the-shelf physical tools with relative ease.

This progression mirrors the evolution of AI coding, where models moved from assisting humans with string-replacement to independently managing software tools. Anthropic posits that the same trajectory is likely occurring in hardware, moving from models being helpful to humans, to humans being helpful to models, and finally to models operating autonomously.