Simon Willison documents Claude Fable 5 proactivity with measured session costs

Category: Agentic coding
Status: discussion
Sources: simonwillison.net, HN discussion
Summary: Simon Willison describes Claude Fable 5 debugging a scrollbar bug in Datasette Agent. Without being asked, the model drove Firefox and Safari, built PyObjC screenshot tooling keyed on window names, generated isolated HTML test pages, injected JavaScript instrumentation that posted measurements to a local HTTP server it wrote, and landed a two-line CSS fix. The session cost $12.11 at full Fable pricing with 68,606 output tokens and a peak context of 113,178 tokens. Willison argues the same proactivity enlarges the prompt injection blast radius and repeats his case for sandboxing coding agents.
Why it matters: Concrete behavior and cost data for Mythos-class agents informs harness design and the sandboxing tradeoff for autonomous debugging.