In August 2025, Anthropic ran an experiment that sounds quirky at first: can Claude help employees with no robotics experience control a robot dog? The answer back then was yes — significantly better than without Claude. Now Anthropic has published Phase 2, and the results are in a different league.
What was tested
The setup: an off-the-shelf robotic quadruped, various tasks including sensor connection, video streaming, lidar integration, manual control via code, path tracking, and ball detection. In Phase 1, two human teams competed — one with Claude, one without.
In Phase 2, Anthropic let Claude Opus 4.7 run solo. No human assistance. The researcher’s only job: plug in the laptop, enter the initial prompt, approve commands.
The results
Claude Opus 4.7 was at least 10 times faster than human teams on every task completed by at least one team last year. On the four tasks both human teams completed, Opus 4.7 was on average 37 times faster than the Claude-less team and 18 times faster than the Claude-assisted team.
To put it concretely: what humans did in over 6 hours, Claude finished in under 10 minutes. It produced nearly ten times less code than the Claude-assisted human team — and was still more successful.
Where Claude still fails
The robot dog was supposed to autonomously retrieve a beach ball. That requires a real-time control loop: detect ball, correct position, adjust robot. Exactly the kind of thing humans with their hands and some practice can do intuitively.
Claude could maneuver the robot behind the ball and give it a push — but the precise closed-loop control for retrieval wasn’t there yet. It’s the same problem the human teams in Phase 1 also struggled with when trying to solve it programmatically.
What this means
Anthropic emphasizes: this isn’t a dedicated robotics project. The improvements are a side effect of general model progress. Claude wasn’t trained on robot control — it simply benefited from general scaling.
We’ve seen this pattern in other domains: first the model helps humans, then humans help the model, then the model does it alone. In cybersecurity, Anthropic is already there. In physical robotics, we’re at the beginning of that trajectory.
Or as Anthropic puts it: we are plausibly entering the early era of physical agentic AI.
Sources: Anthropic Research