Claude pilots a robot dog 20x faster
In Project Fetch phase two, Claude Opus 4.7 ran a robot dog with no human help and finished last year's tasks about 20x faster than the best human team.
Anthropic published phase two of Project Fetch on June 18, a rerun of a 2025 experiment that gave Claude autonomous control of an off-the-shelf robot dog. This time Claude Opus 4.7, working on its own, finished the same navigation and sensing tasks roughly 20 times faster than the best human team managed last August, when those teams used the older Opus 4.1. On the four tasks both human teams completed, Opus 4.7 averaged more than 18 times faster than the team using Claude and more than 37 times faster than the team without it. It wrote 1,045 lines of working code, against the 10,309 the human-plus-Claude team needed.
The setup was deliberately hands-off. A researcher plugged a laptop running Claude Code into the robot, typed the opening prompt, then only approved each command. Anthropic credits the jump to general model scaling, not robotics-specific training. There was still a ceiling: the models could not manage the tight closed-loop control needed to nudge a beach ball, so the literal fetch failed.
What it means for operators
One year is the figure to sit with. A task that beat human teams in 2025 fell to a single unattended model in under ten minutes in 2026, and the gain came from the base model getting better across the board, not from anyone teaching it robotics.
Treat the speed numbers as Anthropic reporting on its own experiment, not an independent benchmark. The honest read is that general agents are closing the gap on messy physical and coding tasks fast, while still failing at the last inch of precise control. That mix of broad competence and sharp blind spots is the same reason to keep checking AI work before you trust it, and why labs are racing to keep autonomous agents under control.