Engineering brief
Stop Panicking About Token Bills. Watch Cost Per Merged PR
At a glance
The brief is enough
- Relevance
- Practical value
- Radar
- Time horizon
- Now
- Warnings
- Vendor pitch
AI spend is up 60x, but Jellyfish's data shows cost per merged PR fell 26% once salary is included. Tokens are about 2.5% of PR cost.
The token bill is not the metric that matters; cost per merged PR, including salary, is what should drive AI budget conversations.
Watch if you're heading into budget season and need a defensible frame for AI spend versus throughput.
Summary
The headline. Jellyfish research, drawn from its engineering workflow platform, argues that AI tokens are still the cheap part of software delivery. Across its customer base, AI spend rose 60x year over year while merged PRs per engineer rose 43%. More importantly, the total cost to produce a merged PR, counting salary and AI spend together, fell 26%. 83% of studied companies saw the same direction.
Why it matters. The framing changes budget conversations. A 60x increase in token spend sounds alarming at board level, but if salary is 97.5% of the cost of a merged PR and AI is 2.5%, the token bill is not the main cost driver. Even the heaviest AI spenders, the top quintile, sit at 8.5% of PR cost. That makes AI cost containment less valuable than throughput gains in most organizations.
The data behind it. Jellyfish grouped companies by token usage. Light users consumed about 60 million tokens per developer per month; heavy users around 2.5 billion, roughly two orders of magnitude more. Heavy spenders improved merged PR throughput by 67% versus about 40% for the other four quintiles, and cut cost per PR by 32% versus about 25%. Within companies, heavy agent weeks cost double the tokens but produced 27% more merged PRs at 18% lower cost per PR.
The quality question. Faster and cheaper PRs did not show a quality collapse. Revert rates were flat enough that Jellyfish says there is no evidence of improvement, but also no worrying sign of degradation. The throughput gain is not from smaller PRs. PRs actually contain more lines of code, which Jellyfish attributes partly to verbose AI output and to developers feeling the bar to create is low.
The trade-off. Diminishing returns exist. The cost per PR improvement curve flattens sharply as token use grows. Yet the top 10% of company-months spent 8x the median tokens and still got 25% lower cost per PR. The conclusion is not to spend without controls, but to stop treating token spend as the primary metric.
What to watch. This is vendor research based on Jellyfish's own customer data, so the sample is not independent and the methodology is not fully public. The numbers are specific enough to challenge the panic narrative, but a cautious leader should validate the cost-per-PR metric on their own data before rewriting budget policy around it.
My take
The cost per merged PR framing is genuinely useful, and I think every engineering leader should steal it for budget season. Counting salary plus AI spend as one bucket changes the conversation with finance, because it gives you a denominator that actually means something.
I'd be careful about the vendor source. Jellyfish sells the tool that tracks this metric, and the methodology is not public, so the exact numbers should be treated as directional, not gospel. Still, the direction is consistent with the obvious math: token bills feel painful, but they are small next to payroll. I would push back on the spend-with-confidence tone. We need a quality counterweight, especially as AI writes more code. Track reverts and incident rates, not just throughput.
What to do with this
- Add a cost per merged PR metric to your monthly review, combining salary and AI spend.
- Ask your platform team to segment token spend by team or quintile before next budget review.
- Set up a revert rate dashboard as a quality counterweight to AI-driven throughput.
Watch the video
This video is blocked due to your privacy settings. To watch this video, please accept YouTube marketing cookies.
Related breakdowns
Agent trust is now an operating problem, not a model problem
An enterprise M&A veteran says agent go-live proves function, not trust. The real test is value, accountability, and evidence before autonomy.
Containment is a vanity metric: agent escalations can be the real work.
Containment is a vanity metric for AI agents. Druid's benchmark shows the same number means different things by industry, and escalation can be revenue.
AI coding assistants are boosting volume — but quality is suffering
DX's platform data from 200,000+ engineers reveals a tradeoff: AI boosts raw throughput but PR size grows 64%, change confidence drops 6%, and change failure…
Get TL;DW
Too Long; Didn't Watch.
A concise breakdowns of the AI and devtools videos that actually matter for engineering leaders.
Free. Weekly. No hype.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.