sqlite-utils 4.0rc2: Claude Fable writes most of the release for $149

Simon Willison, creator of Datasette and co-creator of Django, has published sqlite-utils 4.0rc2, a release candidate that, as he explains on his blog, was mostly written by Claude Fable at an approximate cost of $149.25. The figure is not a minor detail: it is one of the few public cases where a veteran open source maintainer documents how much it costs, in hard cash, to delegate the bulk of a release to a model.

The context goes back two weeks. On June 21, Willison published the first release candidate for 4.0 of sqlite-utils, his library for manipulating SQLite databases from Python and from the command line. Willison follows SemVer with discipline and tries to keep incompatible versions as rare as possible, so he wanted to reach a stable 4.0 with no loose ends. There was also a calendar factor: as he tells it, Claude Fable will only remain available on Anthropic's Max subscriptions for a few more days, so he decided to make the most of it as a final reviewer while he still has it at hand.

A final review launched from an iPhone

The workflow stands out for how mundane it is. Willison kicked off the process from Claude Code for web on his iPhone, with a single prompt asking for a final review before shipping the stable 4.0 and stressing the key point: spotting any last minute issue that would become a breaking change if fixed after the release.

The result was an initial report that Willison has published in full in the project repository. As he tells it, the model found significant problems he had not yet run into on his own, and classified five of them as release blockers: bugs that, if fixed after the stable 4.0, would have forced another compatibility break.

Why it matters

The interesting part of the experiment is not that a model writes code, already routine in 2026, but the role Willison gives it: final reviewer of a major version, the riskiest moment in a library's life cycle. A breaking change that slips into a 4.0 haunts the maintainer for years in the form of issues, patches and users stuck on old versions. That the model caught five blockers the author had not seen, in a library he has maintained for years, says more than any synthetic benchmark.

Transparency is the other half of the value. Willison publishes the full review report in the repository, reproduces the exact prompt he used and puts a price on the work: about $149.25. Against the cost of a flawed major release, measured in support hours and user trust, the math speaks for itself. But that math can only be done if someone publishes the numbers, and that is exactly what almost nobody does.

Who this is useful for

For open source maintainers, the pattern is directly copyable: before a major version, ask the agent for a specific pass hunting for breaking changes, with a short prompt and an auditable report as the outcome. For product teams, the lesson is the same translated to their internal APIs, where a silent break costs more than it seems. And for anyone using sqlite-utils in their projects, the stable 4.0 is close and it is worth going over the breaking notes before upgrading.

From ElephantPink, our reading is measured but clear: the agent as the last filter before a release is one of the best cost-benefit uses we have seen documented this year. It does not replace the maintainer's judgment, who still decides what to fix and what to accept, but $150 for five blockers caught in time is a hard price to argue with.