Project Fetch Phase Two: Claude Opus 4.7 Outpaces Humans on Robot Tasks

Anthropic‘s original Project Fetch showed that a human team with Claude Opus 4.1 could outperform an internet-only team on robotics tasks. Eight months later, the authors revisit the experiment with Claude Opus 4.7 — and the gap has widened dramatically. Opus 4.7, operating without any human assistance, finished the same four tasks roughly 37 times faster than the unaided human team and 18 times faster than the human-plus-Claude team from August 2025. The underlying tension is clear: general-purpose LLM capabilities have improved so rapidly that a model can now effectively control an off-the-shelf robot, yet it still fails at delicate closed-loop manipulation like nudging a beach ball. The progress is not from targeted robotics research but from broad scaling — a pattern that keeps repeating.

The experiment used Claude Code with adaptive thinking and maximum effort. A researcher’s role was minimal: plug in a laptop, enter the initial prompt, approve commands, and advance to the next task. Opus 4.7 quickly chose the correct sensor interfaces and wrote code that was almost always effective on the first try, producing ten times less code than the original Claude-aided human team. Where humans hesitated between multiple approaches, the model committed to a path and executed it reliably. However, the final fetch task — autonomously moving a robot to gently push a beach ball back to home base — remained out of reach. The model could position the robot behind the ball but lacked the precise, reactive control that humans develop through physical practice.

The takeaway for builders: the gap between ‘helpful to humans’ and ‘autonomous in the physical world’ is narrowing faster than most expect. Non-experts can now delegate entire robot-control pipelines to a model, transitioning from pair-programming to oversight. Yet physical tasks requiring subtle, real-time correction still demand human-in-the-loop. The authors suggest we may be entering the early era of physical agentic AI, where models use off-the-shelf tools nearly as fluidly as they now use software editors. For serious technical readers, this means investing in interfaces that let models plug into hardware quickly — and staying alert to the pace at which general scaling erases what seem like fundamental control barriers.