Research reportApril 202614 min read
The Rework Tax
Understanding the real cost of rebuilds and rollbacks
Why AI coding bills ran past annual budgets in early 2026, how much of that spend paid for code that did not last, and what the evidence says about engineering process as the cause.
Open the report, 10 pagesDownload the PDFSummary
Five findings
- 01
Budgets broke early. Uber’s CTO said the company spent its planned 2026 AI coding budget in about four months. In 2025, Anthropic added weekly limits after one $200-a-month subscriber consumed “tens of thousands” of dollars in model usage.
- 02
Spend grows much faster than output. The heaviest token users in Jellyfish’s Q1 2026 sample shipped about twice the merged pull requests of the lightest, at $89.32 per PR against $0.28.
- 03
Much accepted code is later rewritten. Waydev reports initial acceptance of 80 to 90% falling to 10 to 30% once later revisions are counted. Faros AI measured an 861% rise in code churn under high AI adoption.
- 04
Review is the new bottleneck. In the same Faros data, median time in review rose 441.5%, PRs merged with no review rose 31.3%, and incidents per merged PR rose 242.7%.
- 05
The link to architecture is plausible but not proven. No source here measures design scrutiny against token spend directly. Usage incentives, consumption pricing and always-on agents are separately documented drivers of the same bills.
Output rose by tens of percent. Review time, incidents and churn rose by hundreds.
The detail
Seven sections, and what each one rests on
The full report is in the PDF above. These are the pieces worth pulling up on their own.