sqlite-utils 4.0rc2 Cost $149.25 in Claude Fable: A Real AI Coding Cost Case Study

A Rare Public Cost Trace for AI-Assisted Maintenance

Simon Willison's sqlite-utils 4.0rc2 write-up is valuable because it gives numbers that most AI coding case studies omit. The release was built with 37 prompts, 34 commits, 30 files changed, Claude Code using Claude Fable, GPT-5.5 cross-review, and about $149.25 in unsubsidized cost. That makes it a useful real-world benchmark for maintainers asking whether AI coding agents are saving money or simply moving cost from salary time into token bills.

The headline number is not the whole story. A $149.25 model bill can be cheap if it replaces a day of senior-maintainer labor, expensive if it produces review burden, and misleading if the model cost was subsidized by a subscription plan. The useful lesson is the shape of the work: many prompts, many commits, broad but bounded file changes, and a separate review pass using another frontier model.

The Case Study Numbers

The Cost Per Unit Looks Reasonable, But Not Free

Dividing the reported cost across the work gives practical unit economics. $149.25 across 37 prompts is roughly $4.03 per prompt. Across 34 commits, it is roughly $4.39 per commit. Across 30 changed files, it is roughly $4.98 per file. Those numbers are not scary for a maintained open-source project, but they are high enough that repeated agent loops need governance.

The reason the bill can still be rational is labor substitution. If an experienced maintainer values their time at $100 to $200 per hour, the AI cost breaks even quickly if it saves even one to two focused hours. But if the agent generates fragile changes that require a long manual cleanup, the economics invert. The token bill is only one side of the ledger; review time, test time, and defect risk complete the picture.

Why 34 Commits Is the Most Interesting Number

Many teams measure AI coding cost only by prompt count or token spend. The sqlite-utils example shows why commits are a better operational boundary. A commit is reviewable. It can be reverted. It can be compared to test output. It also creates a natural moment to stop the model, inspect the diff, and decide whether the next prompt should continue, narrow scope, or switch models.

That pattern is especially important for agentic tools such as Claude Code. Letting a session run across many unrelated changes can inflate context and hide mistakes. Keeping commits small converts AI work into a series of checkpoints. The cost may still be $149.25, but the risk profile is much better than a single massive generated patch.

Cross-Review Is a Cost, Not a Luxury

The GPT-5.5 cross-review step is a useful reminder that AI coding bills do not end when code is generated. Serious teams should budget for a second pass, either by a human reviewer, a different model, or both. Cross-review adds token spend, but it can lower total cost when it catches regression risk before it reaches manual QA or users.

The mistake is treating review tokens as waste. In a maintenance release, the expensive failure mode is not that the model used too many tokens. It is that the model changed behavior in a subtle way and nobody noticed. For projects with stable APIs, CLI behavior, or database migrations, review cost belongs in the planned budget.

A Budget Template for Similar Maintenance Releases

What Teams Should Copy From This Workflow

• Track prompts, commits, files, and model cost together. Any one metric alone can mislead you.
• Prefer small commits. Agent output becomes cheaper to review when each step has a clear boundary.
• Budget for cross-review. A second model pass is often cheaper than late defect discovery.
• Compare against human hours, not zero. The real question is whether the combined token and review cost beats the manual alternative.

Bottom Line

The sqlite-utils 4.0rc2 example does not prove that every AI-assisted release should cost $149.25. It proves that real agentic maintenance has measurable unit economics. For teams planning their own AI coding budgets, the right takeaway is simple: measure prompt count, commit count, changed files, review passes, and human verification time together. Use the AI Cost Estimator to model the token side, then add the review labor that makes the release safe.