AI NewsProductivityReported
Jellyfish says the median developer used 59M tokens a week
Jellyfish released its AI Engineering Trends report drawn from 276,000 developers. The median developer consumed 59M tokens a week in September, about 8x the level a year ago.
Why it mattersA CFO who approved an AI tools line for 2026 did not budget for 10x growth. Jellyfish says the spend tripled per engineer while headcount barely moved. The question moves from whether to buy, to how to measure what the buy is doing.
A CFO who approved an AI tools line for 2026 did not budget for 10x growth. Jellyfish, a software engineering intelligence company, released its AI Engineering Trends report. The figures are drawn from 276,000 developers at hundreds of companies. TechCrunch covered the report alongside Anthropic's new admin reports for Claude Code. The Jellyfish numbers are the measured part.
The headline figure: the median developer consumed 59 million tokens a week in September. That is about 8x the level a year earlier. Jellyfish measures token consumption by summing tokens reported from Claude Code, Cursor, Codex and GitHub Copilot CLI. It does this for each developer each week. The number is Jellyfish's own measurement across its customer base.
Spend and headcount moved independently
Jellyfish reports AI spend per engineer rose 3.8x from the first quarter of 2026 to the third. Headcount at the median company ended the third quarter at 98 percent of its first-quarter level. The fitted line between the two is flat, meaning the two moved on their own. Jellyfish states this as a correlation across companies, not a causal claim about any one team.
Nicholas Arcolano is the chief scientist at Jellyfish. TechCrunch cites him saying spend on Anthropic-based tools has grown 10x since the start of the year. The cost is now about 2 percent of total engineering cost. Arcolano is quoted saying most CFOs did not plan for 10x growth. He says the question has moved from spending more to making the spend work.
Cost per pull request fell by a third
Jellyfish measures cost per merged pull request as engineering salary plus priced AI usage, divided by human-opened PRs merged. The median company's cost per merged PR fell by about a third over the trailing twelve months. It went from 1,202 dollars to 820 dollars. The spread between cheapest and most expensive companies narrowed. The driver of the fall was throughput: companies merged more PRs per engineer while salary cost held about flat. AI's direct slice is about 18 dollars of the 799 dollar average cost.
Agents open 28 percent of merged PRs at the median
Jellyfish tracks a metric called Autonomous Agent Activity. It is the share of merged pull requests opened or committed by an agent identity. September was the single largest month-over-month step. The median company now sees 28 percent of its merged PRs with agent involvement. Leading companies see 58 percent. Jellyfish also tracks a stricter metric for pull requests whose only reviewer was an AI tool. The median there is 0.5 percent. The leader is 9 percent, and the gap is widening.
Anthropic's new reports, released the same day according to TechCrunch, feed the same question. They show managers what their teams produced, what it cost and where workflows broke, from internal analytics and employee surveys. Anthropic says this data will not be used for model training by default. The company is also adding pooled group budgets so teams can share spend caps.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
