On June 18, 2026, Anthropic’s Frontier Red Team published “Project Fetch: Phase Two.” Researchers Michael Ilie, C. Daniel Freeman, and Kevin K. Troy let Claude Opus 4.7 autonomously drive an off-the-shelf robotic quadruped through the same task set human teams attempted in the August 2025 first phase. The headline result: on tasks humans once completed, the model ran roughly 20x faster than the fastest human team.
This is not another demo video. It is a repeatable, head-to-head experiment on an identical task set — which makes it one of the few concrete data points in the “can LLMs do physical tasks” debate, complete with one very visible failure.
Where the 20x Speedup Comes From
The team ran Opus 4.7 inside Claude Code three times, with adaptive thinking set to maximum effort. On tasks humans completed, the model was about 20x faster than the fastest human team; on every task at least one human team finished, it was at minimum 10x faster. Across the four tasks both sides completed, it beat the Claude-less human team by more than 37x and the team using Claude by more than 18x — while matching or beating their success rate with roughly 10x less code.
That last detail matters. The speedup did not come from brute-forcing with more code; the code volume dropped by an order of magnitude. The post also notes low within-task variance in completion times for finished steps, which for production automation pipelines is often worth more than raw average speed.
How the Experiment Was Run
The design deliberately minimized the human role: the researcher plugged in a laptop, typed the initial prompt, and approved command executions and task transitions. Nothing else. Physical-controller tasks were excluded. Opus 4.7 was chosen because it was the most advanced non-Mythos-class model available; preliminary Mythos Preview trials were not comparable due to setup and serving differences — an exclusion that spawned plenty of speculation once the post hit Hacker News.
Three strengths stood out. The model quickly identified the best sensor-interface approach, the exact spot where human teams had floundered. Most of its code worked on the first attempt. And when it caught itself defaulting to an outdated object detection algorithm, it worked around the mistake on its own.
The Beach Ball It Couldn’t Move
The task that gives the project its name — nudging a beach ball into position — is precisely the part the model could not do. It positioned the robot behind the ball correctly, but the nudging was, in the authors’ words, poorly controlled and unsuccessful. Their diagnosis: the model lacks the closed-loop perception-and-correction skill humans build through practice. For contrast, a more experienced robotics researcher did succeed at autonomous fetching, and the authors believe current Claude likely could too, given more time and scaffolding.
The paper’s bluntest line: “This doesn’t mean that LLMs have now solved robotics. Far from it.” No low-level control such as actuation policies was tested, and the gains came from general scaling, not from any robotics-specific training.
The Same Script, Now in the Physical World
The authors name a pattern: models assist humans, then humans assist models, then models act independently. That script already played out in cybersecurity; Project Fetch documents it emerging in physical-world tasks. We are, they write, “plausibly entering the early era of physical agentic AI.”
For safety research, the alarming variable is not what the model did right but the pace: under a year, zero targeted robotics training, and completion times on the same task set compressed 20-fold. It is also a reminder that longitudinal real-task evaluations and static benchmarks measure different things. Tool-use benchmarks — such as Hugging Face’s agentic tooling evaluation — measure whether a model can use tools at all; Project Fetch measures whether it can converge on a physical outcome alone.
The Pushback, and What It Means for Builders
The Hacker News thread (73 points, 26 comments) split three ways. Skeptics argued the post “mostly reads as a comparison between Opus 4.7 and 4.1,” and that the informative experiment would rerun the human teams with the newer model. Others distrusted first-party research outright. The rebuttal: Opus is publicly available, the hardware is affordable, anyone can replicate it — and a corporate blog is an advertisement in the open, not covert PR.
Three practical takeaways for builders. First, treat LLMs as general-purpose engineers who write control code, not as systems with built-in motor control — closed-loop correction remains a clear weakness. Second, do not overlook low variance: predictable per-step completion times are often what decides whether an automation can ship at all. Third, never take a single lab’s self-reported numbers as your baseline; wait for independent replication.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.