Claude Opus 4.7 takes the leash in Anthropic's robot dog tests

Remember Anthropic's Project Fetch? Back in August 2025, the lab gave a bunch of non-roboticist employees an off-the-shelf robodog and a warehouse to see if Claude could help them fetch a beach ball. The humans struggled, a runaway robot almost plowed into a team, and the flagship model at the time, Claude Opus 4.1, failed completely when asked to run the dog on its own.

Now, Anthropic is back with Project Fetch Phase Two, and the curve has turned sharply vertical.

Cutting code by 90% and running 20x faster

In the new trial, researchers took their hands off the wheel. A human only plugged in a laptop running Claude Code, typed an initial prompt, and clicked approve on commands.

Running with adaptive thinking set to maximum effort, Claude Opus 4.7 operated autonomously across the pipeline. It figured out how to interface with the robodog's video and lidar sensors, wrote custom teleoperation code, tracked the machine's spatial trajectory, and set up computer vision to find the ball.

The results were staggering. Across every task finished by human participants in the original test, Opus 4.7 completed the work roughly 20 times faster than the fastest human team. On the four core tasks finished by both original groups, Opus 4.7 was over 37 times faster than the unassisted human team and 18 times faster than the humans who had Claude Opus 4.1 by their side.

Even more striking was the efficiency. Humans typically spent hours debating sensor protocols and writing bloated boilerplate. Opus 4.7 nailed most scripts on the first try, producing nearly ten times less code than the human-plus-Claude team while achieving the same or better reliability.

The beach ball problem

Despite the speedrun, Claude still flunked the titular part of Project Fetch: actually fetching the ball.

Humans with gamepads can naturally compensate for micro-errors, feeling out how a nudge affects a rolling beach ball and correcting in real time. Opus 4.7 managed to position the robodog squarely behind the ball, but its attempts to gently shepherd the ball back to base were crude and failed, just like the human non-experts attempting the same autonomous script.

Anthropic emphasized that none of this was trained into Claude. There were no robotics-specific fine-tuning runs or bespoke actuation policies. The performance jump was entirely an emergent property of broader language and agentic reasoning scaling.

Why it matters

We think this benchmark highlights a quiet but monumental shift: AI agents are breaking out of software sandboxes and into the physical world.

Anthropic points out a three-stage pattern that already played out in coding and cybersecurity. First, models help humans. Then, humans assist models. Finally, models do the heavy lifting themselves. Hardware tooling appears to be entering that third phase for scaffolding tasks.

For developers and hardware tinkerers, this changes rapid prototyping overnight. Setting up sensor drivers, calibration loops, and bridge code used to be days of grunt work. Claude Code just proved an LLM can automate that plumbing in minutes without specialized robotics knowledge.

However, it also clearly delineates where LLMs currently stop. While Opus 4.7 excels at discrete logical architecture like reading lidar packets, it stumbles on the analog nuance of real-time physical control. Autonomous hardware will likely require a hybrid stack for quite some time, with frontier LLMs orchestrating the high-level brain while specialized physical actuation models handle the feet.