Anthropic: Project Fetch Phase Two Shows 20× Faster Robotic Operation with 10× Less Code
Claude Opus 4.7 autonomously controlled a robotic quadruped and completed tasks 20× faster than a human team while writing ~10× less code with equal or better results, though precise closed-loop control remains a challenge.
This article was generated using artificial intelligence from primary sources.
Anthropic has published results from the second phase of the Project Fetch research initiative, in which Claude Opus 4.7 controlled a commercial robotic quadruped — an autonomous four-legged walking robot — on a set of standardized tasks. The experiment is part of the Frontier Red Team’s work, a specialized group that probes the boundary capabilities and safety limits of the most advanced Claude models.
Results: 20× Speed and 10× Less Code
Opus 4.7 completed tasks approximately 20× faster than the fastest human team working without AI assistance, and around 19× faster than a team using AI tools for support. Across a set of four comparable tasks, the model generated only ~1,045 lines of code, while the human team wrote 10,309 lines to achieve the same results — nearly ten times more. Output quality was equal or higher in favor of the model.
Technical Background
The experiment was designed as a real, unscripted capability test: Opus 4.7 had no pre-defined movement sequences and planned and executed robot commands in real time. The quadruped uses standard commercial specifications — it was not modified for the test — which further validates the applicability of the findings to off-the-shelf hardware. The comparison with the AI-assisted human team (19× slower than Opus 4.7) is key for contextualization: even with tools, teams cannot match the autonomous model in execution speed.
Capability Limits and Safety Aspects
Despite impressive results, the Frontier Red Team identified a clear limitation: precise closed-loop control — the ability to continuously adjust movement based on real-time feedback — still eludes Opus 4.7. Specifically, autonomous ball catching proved reliably unachievable. This limitation is not merely technical but also safety-related: precise physical manipulation is a prerequisite for a range of higher-risk applications. The Frontier Red Team emphasizes that such limitations are an integral part of the research methodology, not accidental oversights.
What This Means for the Industry
Project Fetch Phase Two establishes a measurable reference point for autonomous AI robotics: 20× speedup and 10× code reduction are not anecdotal results but measured across a standardized task set. For industrial applications this means potentially dramatically shorter development and integration cycles. Anthropic has not announced a production deployment date or commercial robotics plans; the findings are treated as research, and Project Fetch remains exclusively within the Frontier Red Team program.
Frequently Asked Questions
- What is Project Fetch and what is its purpose?
- Project Fetch is a research initiative by Anthropic's Frontier Red Team that tests Claude models on autonomous robotic tasks using commercial quadruped robots to measure the potential and safety of AI-driven robotics.
- Why did Claude Opus 4.7 write so much less code?
- Opus 4.7 uses higher-level abstractions and more compact algorithmic patterns. Instead of the 10,309 lines written by the human team, the model achieved equal or better results with only ~1,045 lines, indicating a structurally different approach to problem-solving.