Summary: Anthropic’s June 18, 2026 Project Fetch Phase Two report had Claude Opus 4.7 control an off-the-shelf robot dog through the task set from the August 2025 first phase. On tasks people had completed, the model was reported roughly 20 times faster than the quickest human team. Its one visible failure was a poorly controlled, unsuccessful beach-ball push, showing the gap in closed-loop physical correction.
“Project Fetch: Phase Two” was published on June 18, 2026, by Anthropic’s Frontier Red Team. This time, researchers Michael Ilie, C. Daniel Freeman, and Kevin K. Troy handed Claude Opus 4.7 sole command of an off-the-shelf four-legged robot dog and had it work through the same task set human teams had attempted back in the August 2025 first phase. The result that jumps out: on tasks humans had once completed, the model came in roughly 20x faster than the quickest human team.
This is not one more polished demo clip. It is a head-to-head, rerunnable experiment on an identical task set — one of the few hard data points in the argument over whether LLMs can do physical work, and it arrives with a single highly visible failure.
How the 20x Number Was Produced
Opus 4.7 ran inside Claude Code, three times in all, with adaptive thinking set to maximum effort. On tasks humans had completed, the model landed about 20x faster than the quickest human team; wherever at least one human team had finished, it held at least a 10x advantage. Across the four shared tasks both sides finished, it beat the human team working without Claude by more than 37x and the human team working with Claude by more than 18x — matching or beating their success rate with about 10x less code.
Notice what that means: the speedup was not bought with extra code; the volume of code actually fell by an order of magnitude. The post adds that completion times on finished steps varied very little, and for automation pipelines that need predictability, that property may matter more than average speed.
Running the Experiment With Humans Nearly Out of the Loop
The design was built to shrink the human role as far as possible: a researcher connected a laptop, typed the opening prompt, then approved each command execution and each handoff between tasks — nothing beyond that. Tasks involving physical controllers were left out. Opus 4.7 got the job because it was the most advanced non-Mythos-class model available at the time; early Mythos Preview runs could not be compared fairly owing to differences in setup and serving. That exclusion fed plenty of speculation once the report reached Hacker News.
Three performances stand out. The model locked onto the right sensor-interface approach quickly — the exact spot where every human team had stalled the year before. Most of its code ran on the first try. And when it noticed it had fallen back on the default, outdated object detection algorithm, it detoured around the mistake unaided.
The One Failure: A Beach Ball That Stayed Put
The “fetch” in the project’s name means nudging a beach ball into position, and that is exactly the motion the model never mastered. It placed its body correctly behind the ball, but the push that followed was, in the authors’ words, poorly controlled and unsuccessful. Their reading: humans acquire closed-loop perception and correction through repeated practice, and the model has none of that mileage. The same report holds the counterpoint — an experienced robotics researcher completed the autonomous fetch — and the authors believe today’s Claude could likely manage it too with more time and better scaffolding.
The paper’s least diplomatic sentence reads: “This doesn’t mean that LLMs have now solved robotics. Far from it.” The tests skipped low-level control entirely (actuation policies, for one), and the gains trace back to general scaling rather than any robotics-specific training.
A Script Cybersecurity Already Played Out
The authors spell out a fixed sequence: first models assist humans, then humans assist models, and finally models act on their own. Cybersecurity has already traveled that road; Project Fetch catches the pattern starting over on physical tasks. In their words, we are “plausibly entering the early era of physical agentic AI.”
For safety research, the figure to watch is not what the model got right but the clock: less than a year, no robotics-specific training anywhere in the loop, and completion times on the same task set compressed 20-fold. The study doubles as a reminder that longitudinal evaluations on real tasks and static benchmarks measure different axes. Tool-use benchmarks — Hugging Face’s agentic tooling evaluation is one example — test whether a model can operate tools; Project Fetch tests whether it can carry a physical task to a result on its own.
How the Community Reacted, and the Lessons Builders Should Draw
The Hacker News thread (73 points, 26 comments) settled into roughly three camps. Doubters said the post “mostly reads as a comparison between Opus 4.7 and 4.1,” arguing that the experiment with real information value would put the newer model in the human teams’ hands and run the contest again; some questioned first-party research on principle. The answer from the other side: Opus is available to anyone, the hardware costs little, replication is open to everybody — and a corporate blog is publicity conducted in plain sight, with no disguise attempted.
Three takeaways carry practical weight. Treat the LLM as a versatile engineer who happens to write control code, not as a system with motor control built in — closed-loop correction remains its evident weak spot. Give low variance its due: whether every step lands on a predictable schedule often decides if an automation ships at all. And do not adopt one lab’s self-reported figures as your baseline; hold judgment until an independent replication lands.