On July 4th — while Simon Willison was enjoying the Independence Day parade in Half Moon Bay — Claude Fable was working on sqlite-utils 4.0rc2. 37 prompts, 34 commits, 1,321 new lines of code across 30 files. Estimated API cost: $149.25.
A Critical Bug a Human Would Have Missed
Willison started with a simple prompt: ‘Final review before shipping a stable 4.0 release — very important to spot any last minute things that would be a breaking change if we fix them later.’
Fable found five release blockers. The worst: delete_where() never committed its changes. If you deleted data, inserted new rows, and closed the connection — you lost everything. The deletion, the new data, even new tables. A classic silent data loss bug that’s nearly impossible to catch in normal testing.
Fable as a Coding Partner, Not a Typist
What makes Willison’s writeup so compelling is the working relationship he describes. While Fable spent 10–15 minutes churning through a task, he went to the parade, occasionally checking his iPhone and firing off the next prompt. The heavy lifting ran through Claude Code for web, controlled from his phone.
Fable didn’t just write fixes — it produced the complete documentation for the new transaction model. Willison’s verdict: the release notes are better than he would have written himself.
Cross-Checking with GPT-5.5
Willison then had OpenAI’s GPT-5.5 review Fable’s work — and vice versa. He calls the practice ‘weirdly superstitious’ but admits it works. GPT-5.5 actually caught two more issues with db.query() that Fable had missed.
The Fable Countdown
There’s a time-pressure angle here: starting July 8th, Fable calls will be billed via usage credits even for Max subscribers. Willison upgraded to the $200/month Max plan specifically to squeeze out as much Fable work as possible — he’s currently at 63% of his weekly Fable limit.
What This Means in Practice
$149 for a major release of an established open-source project — including bug fixes, documentation, and a critical security fix. This isn’t a toy demo. It’s a tool thousands of developers use daily.
But Willison himself puts it in perspective: he should have leaned more heavily on cheaper subagents. $141 of the session ran entirely on Fable. A mix of Fable for the hard tasks and Opus for routine work would have brought the cost down.
Sources: