Project Fetch: Phase Two

Source

• Source name: Anthropic Research
• Source kind: blog (frontier-lab research post)
• Authors / organization: Michael Ilie, C. Daniel Freeman, Kevin K. Troy — Anthropic Frontier Red Team
• Venue: Anthropic Research Blog
• Published: 2026-06-18
• Collected: 2026-06-23
• URL: https://www.anthropic.com/research/project-fetch-phase-two
• PDF:
• DOI:

Plain-Language Brief

Anthropic's Frontier Red Team ran a follow-up to their August 2025 "Project Fetch" robotics experiment. Using Claude Opus 4.7 with adaptive thinking (maximum effort) and Claude Code, they measured how fast the model could complete robotics tasks on a quadruped robot compared to human teams from the prior experiment. The model completed tasks 18.9–37.7x faster than the human baselines while writing approximately one-tenth the code.

What the Document Says

• Claude Opus 4.7 completed tasks approximately 37.7x faster than the non-Claude human team and 18.9x faster than the Claude-assisted human team from the August 2025 baseline
• The model generated roughly 10x less code than the human Claude-assisted team while achieving equal or better results
• Researchers characterize this as a progression from "AI assists humans" → "humans assist AI" → "AI operates largely independently"
• The researchers describe this pattern as previously observed in cybersecurity and now appearing in physical robotics
• The experiment omitted low-level robotic control elements (actuation policy development)
• The model still struggled with precise closed-loop physical tasks (autonomous ball retrieval)

Key Concepts Mentioned

• Robotic quadruped (off-the-shelf "robodog")
• Claude Code as scaffolding
• Adaptive thinking / maximum effort mode
• Frontier Red Team
• Autonomous task completion vs. human-in-the-loop

Claims Made by the Document

1. Claude Opus 4.7 completed shared robotics tasks 37.7x faster than unassisted humans and 18.9x faster than Claude-assisted humans
2. The model wrote ~10x less code to achieve equivalent results
3. Performance was reliable across repeated trials with minimal variance
4. The shift from AI-assisted to AI-autonomous operation represents a qualitative capability change

Evidence Used

• Three trials of Claude Opus 4.7 on the same task set as August 2025 experiment
• Comparison against documented completion times from the August 2025 baseline
• Qualitative assessment of task success

Methods / Approach

Replication of August 2025 experiment using Claude Opus 4.7 with adaptive thinking at maximum effort. A researcher handled initial setup and command approval. Completion times and code volume measured across three trials.

Limitations Stated by the Authors

• "This doesn't mean that LLMs have now solved robotics"
• Model struggled with precise ball manipulation (closed-loop physical control)
• Experiment omitted low-level actuation policy development
• No ablation across different model configurations or hardware setups

Limitations Noted by the Agent

Agent interpretation — not from the source.

• Small sample: three trials, single robot, single task set
• Baseline human teams from a separate prior experiment — not directly controlled comparison
• "Frontier Red Team" framing means this is internal Anthropic research; not peer-reviewed
• Speedup metric depends on human baseline quality from prior (different) experiment context
• Industry publication: counts as blog/marketing material even if produced by a research team. No independent replication.

Open Questions

• What low-level robotic control tasks remain unsolvable by current models?
• How does performance degrade as task novelty increases?
• Does this pattern generalize to other off-the-shelf robotic platforms?
• What safety properties were evaluated for autonomous command approval?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• Relevant to questions about AI agent autonomy progression timelines
• Connects to the "AI assistance → AI autonomy" framing used in Anthropic safety research
• [Interpretation] The 10x code reduction claim, if reproducible, has implications for software engineering task complexity metrics

Raw Abstract / Excerpt

"In August 2025, we ran an experiment to see how much Claude could help Anthropic employees—who were not robotics experts—perform sophisticated (and amusing) tasks with an off-the-shelf robotic quadruped (henceforth, a robodog)."