For the past couple of years, “AI costs are exploding” has been a recurring anxiety in engineering budget conversations. Token spend (the amount paid to run AI models, measured in the chunks of text they process) was up 60x year-over-year at many companies by mid-2026. That number, standing alone, sounds alarming. The problem is that it’s the wrong number to be watching.
The unit that actually matters
Developer salary is, by a wide margin, the dominant cost in software engineering. When Jellyfish — whose platform aggregates one of the larger datasets of engineering workflow and AI telemetry data in the industry — looked at what it actually costs to get a merged pull request out the door, the picture shifted considerably. On average across the companies they studied, AI spend accounts for roughly 2.5 percent of the total cost per merged PR. Developer salary is the other 97.5 percent.
Even the heaviest AI users — the top quintile by token consumption — are spending under 10 percent of their PR cost on AI. The idea that AI spend is running away from its value is hard to sustain once you hold it next to the salary line.
The more useful framing, then, is: what is the combined “intelligence cost” to ship one unit of work? Take salary, add AI spend, divide by merged PRs. That’s the number worth tracking.
What the data actually shows
When cost per merged PR is used as the headline metric, the trend runs the other way from the anxiety. Over the past year, despite the 60x growth in AI spend, the dollar cost per merged PR fell by 26 percent across the dataset. Eighty-three percent of companies in the study showed a decline. The mechanism is straightforward: throughput is growing faster than AI costs. More PRs get shipped per engineer, which spreads the salary cost across more work — and that efficiency gain outpaces the rise in token bills by a wide margin.
The relationship between token use and outcomes holds up when companies are broken into groups by usage. The lightest users — around 60 million tokens per developer per month — see roughly a 40 percent gain in PR throughput and a 25 percent drop in cost per PR. The heaviest users — around 2.5 billion tokens per developer per month, about two orders of magnitude more — see a 67 percent throughput gain and a 32 percent cost-per-PR reduction. The curve shows diminishing returns as spend increases, but diminishing returns are not the same as no returns. The top 10 percent of company-months in the dataset use eight times as many tokens as the median but produce PRs that are 25 percent cheaper.
At the individual level, the pattern holds too. Comparing each developer’s heavy-agent-use weeks against their own lower-use weeks: roughly twice the token spend, 27 percent more merged PRs, and each PR about 18 percent cheaper when salary is included.
The quality question
A reasonable counterargument to “spend more, ship more” is that throughput gains might come at a quality cost — faster output that creates more problems downstream. The revert rate (merged PRs that undo earlier work) provides a rough proxy here. In the data, revert rates were flat or slightly declining across all spend groups, including the heaviest. PRs are also not getting smaller — lines of code per PR are actually increasing with AI use, which suggests developers are taking on more scope per ticket rather than submitting thinner work to inflate their numbers.
What this means for how we measure AI investment
The practical implication is not to ignore token costs entirely — runaway background jobs or inefficient prompt patterns are worth catching. But cost-control instincts that operate on token spend in isolation can block investments that are clearly net positive when the full picture is visible. Before implementing token caps, model routing to cheaper models, or usage ceilings, the question worth asking is whether the constraint will improve total cost per PR or just reduce the AI line item while leaving the larger salary cost untouched.
The metrics worth keeping side by side: cost per merged PR (the headline), AI spend per PR, merged PRs per engineer, and a quality counterweight like revert rate. Together they tell a coherent story; any one of them alone tells a misleading one.
Where this is heading
As AI assistants move from autocomplete toward fuller agent workflows — running tests, opening branches, reviewing their own output — token spend per developer will continue to climb. The companies that treat that climb as a red flag risk optimizing for the wrong line. The more durable question is whether the intelligence being purchased is generating more value than it costs, measured in real units of shipped work. So far, for the large majority of the companies in this data, the answer is yes.
Text summarized and optimized using Anthropic’s models and reviewed by a human.