Introducing Claude Opus 5

Summary

Simon Willison's link share (noted as untested because it was written while skiing) covers Anthropic's launch of Claude Opus 5. The model offers near-frontier intelligence at half the price of Fable 5 and currently sits at the top of the Artificial Analysis leaderboard.

Why it matters

The half-price positioning and the model's demonstrated proactivity signal a shift in marketing and behavior. The model autonomously built a computer vision pipeline to solve a task it could not directly perceive. This highlights how frontier models are being marketed and how far agentic behavior has advanced. Additionally, the deliberate distinction between vulnerability discovery and exploitation capabilities underscores the ongoing capability-security tension in deploying frontier AI.

Highlights

• Opus 5 leads the Artificial Analysis leaderboard ahead of Fable 5, while its price remains the same as Opus 4.8.
• In a Frontier-Bench task where viewing an image was impossible, the model wrote its own computer vision pipeline to extract geometry from raw pixels.
• Anthropic deliberately rolled back exploitation training. While vulnerability detection improved, exploitation capability remains 'far behind' Mythos 5.

Critical take

Willison explicitly states he did not test the model. Therefore, the post is merely a link share lacking independent evaluation. The half-price claim is a relative positioning compared to Fable 5, not an absolute cost reduction. Furthermore, a single leaderboard top spot does not guarantee broad capability superiority in real-world workloads.