Anthropic's latest Project Fetch results matter because they show a familiar pattern: first the model helps humans do a task, then the model starts doing meaningful parts of the task alone.
What Changed In Phase Two
On June 18, 2026, Anthropic published Project Fetch: Phase two. The company says Claude Opus 4.7, without human assistance, was about 20 times faster than the fastest human team at the tasks completed in the first experiment.
The setup still uses an off-the-shelf robodog and does not claim to solve low-level control. But Anthropic's own description is important: work that previously needed a Claude-assisted team can now be completed by the model operating on its own.
Why This Matters More Than A Robot Demo
The deeper signal is the progression curve. Anthropic explicitly describes a sequence in which models first help humans, then humans help models, and finally models can largely do the thing themselves. We have seen that in software and cybersecurity already. Anthropic is arguing that the same pattern is beginning to appear in the physical world.
For zero-human companies, that does not mean everyone becomes a robotics operator. It means the set of tasks that can be handed from planning systems into embodied execution keeps expanding.
The Limits Matter Too
Anthropic is careful to say the model still struggled with precise ball retrieval and that this work does not address the hardest actuation problems. That caution is useful. The point is not that robotics is solved. The point is that the slope of improvement is steeper than most teams can afford to ignore.
Once models can reliably bridge more of the gap between perception, tool use, and physical action, the architecture questions start to look much more like mainstream agent design.
The Take
Project Fetch phase two suggests physical autonomy is starting to move through the same adoption curve we already see in digital work: assistive, then semi-autonomous, then increasingly self-directed.
That is a real capability signal for zero-human company builders, even if the first major applications stay narrow.
Related: See our previous research on Qwen-Robot Suite, Qwen-RobotNav, and NVIDIA physical AI skills.