Claude Opus 4.7 finishes a robodog coding task about 20x faster

Agent benchmarks are beginning to leave the IDE and touch hardware. Anthropic’s Project Fetch phase 2 tested whether Claude Opus 4.7 could program a quadruped robot dog, and the company says the model working on its own was about 20 times faster than last year’s best human team aided by Opus 4.1.

In the source tweet, Anthropic wrote that Opus 4.7 was "~20x faster" than the prior human-plus-Claude team. The same post included the grounding caveat: the robodog still failed to fetch the beach ball. That makes the result more useful, not less, because it separates rapid code generation and hardware orchestration from full real-world task success.

Project Fetch comes from Anthropic’s Frontier Red Team work. The setup asks people without robotics expertise, and now a model acting more autonomously, to make unfamiliar robot hardware complete a simple physical objective. This is exactly the kind of test that exposes the gap between software fluency and embodied reliability: documentation has to become control code, control code has to survive noisy sensors, and the final behavior has to work in a room rather than a prompt.

The next metric to watch is not only whether Claude can eventually make the robot fetch the ball. More telling measures include the number of intervention points, the repeatability across robot platforms, the safety envelope around autonomous hardware control, and whether speed gains come with brittle assumptions. If agentic models keep improving here, robotics teams may use them first as fast integration engineers before trusting them as full autonomy systems.