How Simon Willison used Claude Fable 5 to ship an SQLite update for $149
An AI agent just did a senior dev's job for the price of a decent dinner, while Simon Willison was eating hot dogs at a parade. This isn't just another autocomplete wrapper; we are officially entering the era of autonomous SQLite codebases.
Willison started with a single prompt from his iPhone using Claude Code, asking the model to do a final pre-release audit of sqlite-utils 4.0, essentially treating one of the world's most advanced AI models like a highly-stressed intern.
The AI agent executed 37 prompts and 34 commits, altering 30 files with over 1,300 new lines of code, while its human creator literally walked away to enjoy a holiday parade in Half Moon Bay, occasionally checking the progress on his phone.
Fable didn't just refactor style; it uncovered five critical bugs that would have broken the release, including a silent database killer where the delete_where method left connections open and silently rolled back all subsequent writes on close — the kind of bug that makes developers question their life choices after a database silently vaporizes in production.
To double-check the work, Willison ran a cross-review using OpenAI's GPT-5.5 via Codex, which spotted two additional high-priority bugs that Fable then successfully patched, proving that peer review is still alive, even if the peers are just silicon chips arguing with each other.
The entire session cost $149.25 in API credits, calculated via AgentsView, which Willison bypassed by burning through his $200 monthly subscription right before the upcoming "fablepocalypse" pricing shift on July 7.
The days of human developers sweating over edge cases are numbered when a $150 API run replaces a week of manual QA. The real test begins when the cheap subscription loops close, and companies have to decide if they prefer paying junior devs or Anthropic's cloud bill.
Source: Simon Willison's Weblog
Comments
Help shape the next version: Add context or suggest a correction. AI review can add points toward a rewrite. Reviews and updates may take time; a full meter does not guarantee a new version.