Simon Willison shipped a release mostly written by AI for $149.25 — the number everyone will misuse

What $149.25 actually bought

Simon Willison describes shipping sqlite-utils 4.0rc2 mostly written by Claude Fable, then using Fable to run a pre-release review that caught bugs he'd missed. The tally: 37 prompts, 34 commits, +1,321/−190 across 30 files, and five "release blocker" bugs — the worst being delete_where() leaving a connection in an open transaction that silently rolled back later writes on close. He estimates the unsubsidized API cost at $149.25, though in practice it was covered by his $200/mo Claude Max plan. He also ran a cross-model review with GPT-5.5, which caught two more commit-timing bugs Fable had missed — leading him to conclude that having one model review another "really does work," even though it first felt superstitious. He was, by his own account, prompting from his phone at a Half Moon Bay Fourth of July parade.

The thread took the money number apart before anyone else could weaponize it.

The caveat the headline will lose

"$150 spent by an expert, a software engineer with significant practical knowledge in AI, is not equivalent to the exact same amount spent by a novice. … Expect alarmist posts tomorrow morning … claiming building software is now as cheap as diner at the restaurant." — a developer who self-hosts his own infrastructure

A practitioner named the failure mode that makes the novice's bill balloon — these tools are sycophants with a linter:

"the models are still too eager to please. If I ask it to scan a release and note possible issues, it absolutely will find issues. … There is never a "Nothing found, good to go!"." — a developer who runs repeated LLM PR reviews

And a pointed engineering objection: maybe don't ask an AI what's broken — write the test that would have caught it:

"I'm kind of surprised that there is no test case that would have identified the fact that delete_where() leaves the state corrupted. … maybe instead of asking Claude what is wrong it would be wiser to invest in test coverage." — a commenter arguing from testing discipline

Even a peer tool-builder chimed in from experience — Jeremy Howard of fast.ai noted his own apsw port has quietly run in production for years, a reminder the ecosystem around this library is deeper than one release.

Why the number is a trap

Willison's result is real and impressive: an AI caught five shipping bugs in his own code, and a second model caught two the first one missed. But "$149.25 to build a release" is about to become "AI builds software for the price of dinner," and the thread inoculates against exactly that. The cost isn't the API bill — it's knowing which five bugs are worth chasing and ignoring the fifty the model invents to please you. That expertise is the product. The invoice is just what it costs him.