Claude Opus 4.7 Project Fetch: Autonomous API Tasks 20x Faster

What Is Project Fetch Phase Two?

On June 18, 2026, Anthropic's Frontier Red Team (Michael Ilie, C. Daniel Freeman, Kevin K. Troy) published Project Fetch Phase Two — a repeat of an August 2025 experiment where teams used Claude to control a robotic quadruped. The twist: this time, Claude Opus 4.7 ran completely autonomously via Claude Code.

The human researcher's role was minimal: plug in the laptop, enter the initial prompt, approve commands, and approve transitions between tasks. Everything else — code writing, hardware connection, sensor integration, iteration — was handled by Claude.

Benchmark Results: The Numbers

The model was "as or more successful than both human teams while producing almost ten times less code than Team Claude" — meaning its solutions were both faster and more efficient.

What This Means for API Developers

Claude Code (Anthropic's terminal-based agent CLI) is the interface for autonomous task execution. Any developer can install the CLI, launch autonomous mode with Opus 4.7 using adaptive thinking at max effort, describe the task in natural language, and approve critical transitions.

import anthropic

client = anthropic.Anthropic(api_key="YOUR_API_KEY")

# Claude handles multi-step tasks autonomously
response = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=64000,
    messages=[
        {"role": "user", "content": "Write a Python program connecting to a USB camera, capturing 30fps frames, detecting red objects with OpenCV, with keyboard arrow control."}
    ]
)

print(response.content[0].text)

# Claude Code CLI — terminal-based autonomous agent
# npx @anthropic-ai/claude-code --model claude-opus-4-7

# In the CLI, just describe the goal:
# "Connect to USB camera at /dev/video0, capture frames,
#  detect red objects, and show a real-time tracking window."

Pricing: Opus 4.7 vs Competitors

Why This Matters: The Early Era of Physical Agentic AI

The key insight: this wasn't a software-only benchmark. Claude controlled real hardware — a physical robot with cameras, lidar sensors, and motors. Most autonomous coding evaluations (SWE-bench, HumanEval) test pure software tasks. Project Fetch Phase Two demonstrates that Claude can bridge code generation and physical-world interaction.

For API developers: autonomous debugging, end-to-end builds, and reduced human overhead are now practical. What required 3-6 highly skilled engineers working 6 hours can be done by one developer with Claude Code in 10-30 minutes.

Limitations

• Physical precision: Fine motor control (the actual "fetching" — moving a beach ball) still challenged the model
• Tokenizer change: ~30% token increase vs pre-4.7 models requires cost recalculation
• Opus 4.7 is now previous-gen: Opus 4.8 (May 28) and Fable 5 (June 9) are newer
• Export control: Anthropic models face US restrictions affecting international teams
• Single-provider: Claude Code works only with Anthropic models — no built-in multi-provider routing

Autonomous Coding Tools Comparison

Project Fetch Phase Two tests a scenario no other coding tool has benchmarked: hardware-in-the-loop autonomous development. While Devin and Codex focus on software repositories and CI/CD, Claude demonstrates end-to-end hardware interaction.

Frequently Asked Questions

Verdict

Project Fetch Phase Two is a landmark demonstration of what autonomous API agents can achieve. Claude Opus 4.7 completing hardware software engineering tasks 20x faster than humans is not theoretical — it's a replicable result with real code, real sensors, and real robots.

For API developers, the takeaway is clear: autonomous multi-step task execution via Claude Code + Opus 4.7 is production-ready. The model writes code that works on first try, tests it against real hardware, and iterates until the task is complete.