Simon Willison — 2026-07-24#

Highlight#

Anthropic has just released Claude Opus 5, a model promising near-frontier intelligence at half the price of Fable 5. Simon highlights its highly proactive capabilities and nuanced approach to cybersecurity training, noting that it is already topping the Artificial Analysis leaderboards.

Posts#

Introducing Claude Opus 5 · Source Simon notes that Opus 5 is currently leading the Artificial Analysis leaderboard, even beating out Fable 5, while maintaining the same base pricing as the older Opus 4.8. He is particularly intrigued by a benchmark anecdote demonstrating the model’s “relentless proactivity” — when denied access to a drawing for a 3D FreeCAD task, the model autonomously wrote its own computer vision pipeline to extract the necessary geometry from raw pixels. Anthropic also deliberately limited the model’s ability to exploit cybersecurity vulnerabilities while improving its ability to find them, which Simon jokingly hopes will keep the US government from shutting it down. Early experiments with prompt-driven SVG generation are already underway, and Anthropic has released a dedicated prompting guide to help developers get started.

Project Pulse#

While today’s single post is a quick reaction to a major model release, it continues Simon’s long-standing habit of immediately evaluating new LLMs for agentic behaviors and stress-testing their practical capabilities with small developer tasks like SVG generation.