← Back to feed

Anthropic published the results of Project Fetch Phase 2 today: Claude Opus 4.7, running autonomously in Claude Code with maximum adaptive thinking, outperformed both human teams from the original August 2025 experiment across every task they completed — by a factor of 20 on average.

The original experiment staged two groups of non-robotics Anthropic employees against a commercial robotic quadruped. One team had access to Claude Opus 4.1. One did not. Neither was expert. The goal was six tasks: connect to the robot’s sensors, write a manual control program, monitor its path through space, detect a beach ball, and finally combine those into an autonomous retrieval system. The Claude-enabled team finished more and faster.

Phase 2 removes the humans from the loop entirely.

The Numbers

On the four tasks completed by both human teams in Phase 1, Opus 4.7 ran:

• 37x faster than the team without Claude
• 18x faster than the team with Claude

Measured against the fastest human team across all completed tasks, Opus 4.7 averaged 20x quicker. It also produced roughly 10x less code than Team Claude while achieving the same or better outcomes.

The researcher’s role was reduced to plugging a laptop into the robot, entering the initial prompt, and approving command execution and task transitions. No steering. No debugging.

Where It Still Fails

The one task where Opus 4.7 did not match humans: the “fetch” in Project Fetch. Precisely nudging the beach ball back to a home-base patch requires continuous visual feedback and fine-grained motor adjustment — a closed control loop that humans handle intuitively after a few mistakes.

Opus 4.7 could position the robot behind the ball. It could not reliably complete the retrieval with adequate precision. Anthropic notes that a researcher with more robotics experience than the original volunteers did solve this task with additional scaffolding, suggesting current models are close to capable but not there yet without extra structure.

General Scaling, Not Robotics R&D

Anthropic is explicit that these gains are not from targeted robotics training. The improvements emerged from the same general scaling that produced capability gains in coding and reasoning. No robotics-specific dataset, no fine-tuning, no simulation curriculum.

The paper mentions a preliminary attempt with Claude Mythos Preview that was excluded — the experimental setup and how the model was being served at the time prevented a fair comparison. Anthropic says Mythos results will be published separately.

The Pattern Anthropic Is Watching

The team frames Phase 2 as the third stage of a recurring dynamic they have observed in both cybersecurity and now robotics:

1. Models help humans do tasks they could not do alone
2. Humans help models accomplish tasks models could not do alone
3. Models accomplish tasks independently, faster than the humans who trained on them

Project Fetch Phase 1 was stage one. Phase 2 is stage three, at least for the engineering and programming tasks the robot requires. The physical motor control problem — stage three for retrieval — remains open.

What This Means for Physical AI

The commercial implications are narrow for now. The robodog tasks in this experiment are essentially software engineering: reading APIs, writing sensor interfaces, debugging control programs. Those are exactly the tasks at which Opus 4.7 already excels in any other context.

What changes is the loop. The AI is no longer writing code that a human then runs. It is writing code, verifying it executes on physical hardware, observing the output, and adjusting — in a tight, near-unattended cycle. That is a qualitatively different deployment mode from an IDE agent.

Anthropic estimates that current Claude models are close to solving the precision retrieval problem with more time and additional scaffolding. The Phase 3 research question is whether Mythos-class models can close that last gap with the same unattended reliability seen here on the engineering tasks.