# Tokens Are Cheap, But Is the Math Real? _Jellyfish analysis says merged PR cost fell 26% despite 60x AI spend, but quality and causality questions remain._ **Published:** 2026-10-19 **Source:** https://www.startuphub.ai/ai-news/market-research/2026/tokens-are-cheap-but-is-the-math-real --- The cost of a merged pull request fell 26% in a year even as AI spend rose 60x, [AI Engineer](https://www.youtube.com/watch?v=U7ZYBCcdJPg) host Steve Jamison said Oct. 14 in a talk drawn from Jellyfish platform data. The math only works when you combine human and machine costs, Jamison argued. He defined intelligence cost per PR as salary plus tokens, and said tokens were about 2.5% of that cost in August 2026, with salary at 97.5%, so throughput spreading salary thinner outran the token bill. Merged PRs per engineer rose 43% over the same period, he said, and 83% of companies studied saw cost per PR fall. That average hides a spread. Jamison split companies into fifths by tokens per dev per month, from about 60 million tokens at 0.3% of PR cost to 2.5 billion tokens at 8.5% of PR cost. The heaviest fifth saw 67% throughput gains and a 32% drop in cost per PR, he said, versus roughly 40% throughput gains and about a 25% cost drop for the other four fifths. At the individual level, heavy agent-skill weeks cost about 2x in tokens but yielded 27% more merged PRs at about 18% lower cost per PR, per the same analysis. The skeptic read is what the data cannot show. Revert rates, his proxy for quality, ticked down slightly for every group but sat below statistical noise, Jamison said. Lines of code per PR rose, which he attributed to verbose AI output and larger scopes, not smaller tickets. That leaves the core question of causality untouched. Jellyfish customers who spend more on AI are not a random sample, and the dataset does not control for team size, model mix, pricing shifts, or what those extra lines actually ship. Diminishing returns are in the chart, not the headline. Jamison showed cost per PR falling steeply with early token use then flattening, and noted the top 10% of company-months used 8x the tokens of the median for about 25% lower cost per PR. The curve still slopes down in his sample, but he did not price a ceiling, name a budget where tokens stop being the cheap part, or share audited salary assumptions. A controlled comparison of similar teams with assigned token budgets would change the read. Until then the open question is not whether tokens look cheap against salary, but at what volume they stop buying cheaper delivery. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.