News

What $149 of AI Labor Reveals About Developer Value

Developer and tool-maker Simon Willison published a release candidate for sqlite-utils 4.0 on July 5, 2026, with an unusual disclosure: the update was "mostly written by Claude Fable" at an estimated cost of $149.25 across 37 prompts and 34 commits. Willison calculated that figure by running the session through the AgentsView tool to model what the API calls would have cost at standard rates, even though he pays a flat $200-per-month Claude Max subscription.

sqlite-utils is a widely used open-source Python library for working with SQLite databases. The 4.0 release introduces breaking changes serious enough to warrant careful pre-ship review — which is exactly how Willison put Fable to work.

Context

The workflow Willison describes is worth examining closely because it illustrates where human judgment entered and where it didn't. Fable's initial pass flagged five release blockers, one of which — a delete_where() function that never commits, silently poisoning the database connection — was a genuine data-loss bug Willison had not caught. His reaction, quoted directly: "That's a really bad bug! Very glad I didn't ship that."

But Willison didn't stop at one model. He then fed Fable's changes to GPT-5.5 xhigh for an adversarial review. That second pass found two additional priority-1 issues: db.query() committing writes before raising an error on non-row statements, and INSERT ... RETURNING calls committing only after a generator was fully exhausted — contradicting the library's own documentation. On the workflow of using one AI to check another, Willison wrote: "The problem is _it really does work_."

The reaction in the Hacker News thread surfaced a tension that hiring managers and candidates alike are navigating. Commenters noted that an expert's $150 of AI spend "is not equivalent to the exact same amount spent by a novice." The cost figure obscures how much professional judgment was required to select the right tasks, write precise specifications, and know when an agent's output was plausible versus subtly wrong. Others flagged that AI systems struggle to know when to stop — models can enter loops finding or inventing issues — requiring explicit human constraints to get useful output.

What This Means for Job Seekers

Reviewing this workflow, what stands out is that none of the high-value steps were automated: deciding which library behavior constituted a blocker, writing prompts tight enough to produce correct commit-ready code, recognizing a data-loss bug when a model flagged it, and choosing to run a second model as adversarial reviewer. Those are judgment calls. The $149.25 covered execution; a veteran open-source maintainer covered everything else.

That reframing has direct implications for how developers should present themselves right now. If you are preparing for technical interviews or reworking your resume, the skill worth demonstrating is not that you can write code faster with AI — it is that you can specify problems precisely enough for an agent to solve them, and audit the output well enough to catch a silent data-loss bug before it ships. Those are the same skills that distinguish a senior engineer from a junior one, restated for an agentic workflow.

The harder question the Willison case raises is about the job search landscape for developers: as AI agents absorb more of the execution layer, employers will increasingly test for the supervision layer. That means code review, QA judgment, and the ability to write specifications that are unambiguous — not because those skills are new, but because they are now the primary proof of value that an agent cannot yet replicate.