- Steve Jamison of Jellyfish said that in the firm's August 2026 data, AI was about 2.5% of a merged pull request's cost and salary the rest.
- He urged teams to track total cost per merged PR beside the AI bill, throughput and quality, rather than fight over token spend alone.
Steve Jamison of Jellyfish argued on the AI Engineer show that the AI bill is the wrong number to fight over. In Jellyfish's August 2026 data, salary was about 97.5% of the cost of a merged pull request. That is a code change accepted into the main codebase. Capping tokens, the units AI models bill by, aims at the small part.
What was said
Jamison said the two figures he had just shown leave out developer salary. Jellyfish instead puts pay and AI spend into one cost bucket per PR. Overall, AI was about 2.5% of that bucket. He stressed these are relative costs, not dollars.
The share varied widely. Jellyfish ranked companies by token use. The lightest fifth used about 60 million tokens per developer per month. There, AI was about 0.3% of a PR's cost. The heaviest fifth used about 2.5 billion tokens. There, AI was 8.5%.
The mechanism is division. Cost per PR is pay plus AI spend, divided by merged PRs. More PRs spread the same salary thinner. Over the last year, a merged PR cost 26% less in dollars. AI costs rose 60 times in that time. Heavy spenders merged 67% more PRs per developer. Their peers gained about 40%. Jamison said throughput, not AI spend, drives most of the change. On tracking spend, he said: "I'm saying don't let that be your driving metric."
Why it matters
Our reading: a token cap is easy to set and easy to defend in a budget meeting. If Jamison's numbers hold at your firm, a cap trims a sliver of a PR's cost. It may also squeeze the throughput that moves the number most. That link is a hypothesis.
Three budget questions follow. What share of your cost per PR is AI today? Does a cap lower merged PRs per developer? Are reverts and quality metrics holding? Jamison advised tracking total human-plus-AI cost per PR each month. He said to read it beside the AI bill and throughput.
The other side
The show notes call these observational comparisons. They do not prove that token spending caused the throughput gains. They also do not prove that companies realized equal cash savings.
Jamison conceded limits himself. Gains shrink as usage rises. But he said "diminishing returns is not the same as no returns." He said a time will come when tokens are not the cheap part. He added that most companies are not close.
One gap remains. Cost per merged PR counts code that cleared review. It says little about who pays when faulty code ships.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Tokens Are the Cheap Part — Steve Jamison, Jellyfish (2026-10-13)
