Project Fetch: Phase Two

What happened

Anthropic's Frontier Red Team released Phase Two of Project Fetch. An autonomous Claude Opus 4.7 programmed a Unitree quadruped on the same tasks human teams tackled a year earlier with Opus 4.1.

Context and impact

This is internal red-teaming, not a product demo – the team studies how frontier models behave when driving physical hardware. The pattern ("models first assist humans, then humans assist models, then models work alone") now shows up in robotics too, with implications for agentic autonomy and safety arguments.

Details

• For every step at least one human team completed last year, Opus 4.7 finished that step at least 10× faster.
• On aggregate tasks it was ~20× faster than the fastest Phase One human team.
• Opus 4.7 produced almost 10× less code than last year's "Team Claude".
• Weak spot: fine motor control – the model could aim the robot at the ball, but the final push to the target still failed.