The takeaway: Simon Willison — creator of Datasette, co-creator of Django — shipped a release candidate of his widely used sqlite-utils library this weekend "mostly written by Claude Fable", and published the receipt: $149.25 of estimated API-equivalent cost, 37 prompts, 34 commits. He's racing what he calls "the July 7th Fablepocalypse" — the same credits cliff we've been covering — and his write-up is the best ground-level picture yet of what that deadline means for the people who build with these models.

The release, and the bug that justified it

Willison published the account on his blog today: with Claude Fable 5 available on Max subscriptions "for a few more days", he set it loose on the final review of sqlite-utils 4.0 — a stable, semver-major release of a library a lot of Python data work quietly depends on.

The opening 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" — turned up five problems Fable classed as release blockers. The worst was genuinely nasty: a delete_where() call that never committed and left the connection poisoned, so subsequent writes — including entire new tables — silently vanished when the connection closed. "That's a really bad bug!" Willison writes. "Very glad I didn't ship that."

From there: 37 prompts, 34 commits, +1,321/−190 lines across 30 files. Much of it was driven from his iPhone — he went out to the Half Moon Bay 4th of July parade while the agent churned, checking in from his phone between tasks.

The receipt

The number in the headline is unusually well documented. Willison ran a cost-analysis tool inside the session and published the breakdown: $141.02 for the main Claude Fable session, four Fable sub-agent reviews at between $1.40 and $2.40 each, and one 32-cent Opus 4.8 agent — $149.25 in estimated unsubsidised API cost, absorbed by his subscription. He upgraded from the $100 to the $200 Claude Max plan specifically to enlarge his Fable allowance before 7 July, when — as we covered this week — bundled subscriber access to Fable 5 ends and the model moves to paid credits.

His stated plan: "hitting 100% on that Fable bar just in time for the price increase." It's the metered-AI economics working exactly as designed — a flat-rate window closing, and a sophisticated user extracting maximum value on the way out. His one regret cuts the other way: he "really should have" delegated more to sub-agents running cheaper models — the same Opus-versus-Sonnet calculus every Claude user is now doing.

Rival models reviewing each other's work

The methodological detail worth stealing: before shipping, Willison had OpenAI's GPT-5.5 review Claude's work. "I used to think that the idea of having one model review the work of another was somewhat absurd — it felt weirdly superstitious," he writes. "The problem is it really does work." The cross-review found two priority-one issues in the transaction handling — both confirmed by experiment, both fixed before release. He now does this habitually, in both directions.

He also let Fable write the release notes, updating the changelog as each change landed: "honestly these are better than I would have created myself… they need to be boring, predictable and accurate."

Why this one matters

There's no shortage of "AI wrote my code" posts. This one is different because of who wrote it and what it contains: a maintainer with two decades of open-source credibility, a library with real users, a data-loss bug caught before it shipped, a full published transcript, and a line-item bill. It's the most concrete public data point yet on what frontier-model coding actually costs — and a preview of the accounting every developer on a Claude subscription starts doing for real on Tuesday.

Disclosure: On The Wire runs on Anthropic models — including Claude Fable 5. We flag it every time we cover the company.