Two growth rates cannot settle the budget
AI spending is up 60-fold, while merged pull requests per engineer are up 43%. Steve Jamison opens with those two year-over-year changes in Jellyfish’s engineering-workflow and AI-telemetry data. Neither tells an engineering leader whether the next dollar of AI spending is worthwhile.
The finance team can see a much larger bill. The engineering team can point to more merged work. Comparing the percentages is tempting, but they measure different things, and the spending figure leaves out the largest cost involved in producing the code: developer salaries.
Capping token use might stop waste. It might also remove a tool that saves expensive human time. Another company’s bill offers little guidance if its team, workload, and results differ from yours.
Jamison proposes putting both kinds of spending into the same calculation. For a reporting period, add developer salary costs and AI costs, then divide by the number of merged pull requests. He calls this the intelligence cost of a merged PR. The question becomes whether the combined cost of delivering a unit of reviewed code is rising or falling. The reframing begins at 4:46:44.
Count the human work as well as the tokens
In the August 2026 breakdown Jamison presents, AI accounts for 2.5% of that combined cost. Human salary accounts for the other 97.5%. A dramatic increase in the small component can coexist with a reduction in the total cost per PR if the same salaried team merges more work.
At 4:47:35, the relative sizes explain why the token bill alone can mislead. The percentages describe Jellyfish’s presented data, not every engineering organization.
The calculation uses merged PRs, rather than PRs merely opened. That includes the time spent getting work through review and at least some quality gates. It does not turn PR count into a direct measure of customer value, but it gets closer to delivered work than counting generated code alone.
Across the study’s one-year comparison, the reported cost per merged PR fell 26%, even with the 60-fold increase in AI costs. Jamison says 83% of the companies studied saw a decline. His explanation is that throughput increased fast enough to spread salary across more merged changes, outweighing the added AI spending.
The useful comparison is therefore a ratio with consistent scope: combined salary and AI costs divided by merged PRs for the same team and period. Keep the bill and throughput alongside it so a change in the ratio remains explainable.
Heavy users improved more, with diminishing returns
Jellyfish divides companies into five groups by token use. In the lightest fifth, AI represents about 0.3% of the cost per PR; in the heaviest fifth, it represents 8.5%. Jamison reports roughly 60 million tokens per developer per month at the light end and 2.5 billion at the heavy end.
All five groups increased throughput. The lower four groups improved by around 40%, while the heaviest group increased merged PRs per developer by 67%. The corresponding cost per PR fell by roughly a quarter in the lower groups and 32% in the heaviest group.
The comparison at 4:50:25 shows a 23–27% decline for the other four fifths and 32% for the heaviest fifth. The groups differ in token use; this observational comparison does not isolate spending from every other difference between their workflows.
Jamison also shows diminishing returns. Initial increases in token use accompany a steep reduction in cost per PR; the curve then flattens. A flatter curve still leaves room for improvement. In the comparison he presents, the top tenth of company-months by token use consumed eight times as many tokens as the median and had a 25% lower cost per PR.
That does not imply unlimited spending is efficient. It suggests that the point where additional tokens stop paying for themselves had not been reached by much of the observed population. Whether a particular team is near that point requires its own output and cost data.
A developer-level analysis provides another view. Comparing heavier agent-skill weeks with the same developers’ weeks without skills, Jamison reports about twice the token spending, 27% more merged PRs, and an 18% lower combined cost per PR. The displayed sample spans September 2025 through July 2026 and about 690,000 developer-weeks. Comparing people with themselves reduces some differences between developers; it still leaves questions about how their tasks and workflows changed between weeks.
Check whether the output got worse—or just smaller
Two objections matter before treating cheaper PRs as a win. The team might be merging more mistakes. Or it might be dividing the same work into smaller PRs.
For the first, Jamison examines reverts: merged changes that undo an earlier change. The heaviest token-use group started with the lowest revert rate, before much of the token growth. Rates later declined slightly across the groups, but he explicitly says the changes are too small relative to the noise to count as evidence of improvement.
At 4:54:25, the chart uses a three-month average from March 2025 through August 2026. Jamison reads it as an absence of a concerning rise in reverts, not proof that code quality improved.
A revert is only one quality signal. Problems can survive without a rollback, and a PR can pass review without being valuable. Jamison’s operational recommendation is to keep quality measures next to cost and throughput, so a favorable ratio cannot hide damage elsewhere.
The smaller-PR explanation also does not fit the presented data: lines of code per PR increased. Jamison offers several possible reasons, including generated verbosity, unnecessary tests, and developers expanding a change because another edge case is only a prompt away. He does not settle which explanation dominates.
The increase rules out smaller PRs as the simple explanation for this throughput pattern. It also raises its own question about review load. More code is not automatically more useful work.
Budget against delivery cost
Return to the original comparison: 60 times the AI spend and 43% more output. Add the salary cost, and the result in Jellyfish’s study is a 26% lower cost per merged PR. That is the number Jamison wants engineering leaders to bring into a budget discussion.
Keep three views together each month:
- The AI bill: how much the team spends, including runaway jobs that should be stopped.
- Merged PRs per engineer: how delivery throughput changes as the workflow changes.
- Combined cost per merged PR: whether the extra spending is outweighed by lower human cost per delivered change, with quality measures beside it.
This framing also changes how to evaluate optimization projects. A model router, a custom post-trained model, or a token limit might reduce the bill. Helping engineers use agents more effectively might improve throughput enough to save more overall. The comparison needs both sides of the cost equation.
Jamison’s conclusion for the range of organizations he observes is to keep investing while measuring the return. The token bill tells the team what it bought. The combined delivery cost helps it judge whether that purchase made the work cheaper. Watch the budgeting discussion at 4:58:48.