Cursor Developer Habits Report: AI Gains Are Concentrated Among Power Users (Category AI4SDLC)

I have been looking through the spring 2026 Cursor Developer Habits Report. It has interesting charts on development speed, growing context and automation. One of its most interesting insights is that productivity gains are not spread evenly: they strongly favour people who already know how to fit agent work into their day. The study uses aggregated product and engineering data: agent use, tokens, accepted AI changes and merged pull request activity. It excludes data from privacy-mode users who opted out.

The report opens with these findings:

1️⃣ Faster development Code-writing speed doubled year over year, pull requests are getting larger and deeper, and agent-generated code is increasingly passing review. 2️⃣ The economics of intelligence The authors compare seven model families by cost per line of code and cost per task submission, finding substantial differences in unit economics. At a fixed task success rate, Opus is the most expensive, followed by ChatGPT and then Cursor’s Composer 2.5. The gap between Cursor’s model and Opus is orders of magnitude. 3️⃣ The power user gap Although AI produces broad productivity gains, the largest changes occur among the top 1% of developers. 4️⃣ Context matters more Input token counts have risen sharply, with a shift towards cache-read tokens. This gives agents a kind of working memory for harder tasks and better code. 5️⃣ The shift towards automation The authors show coding agents evolving from individual developers’ tools into systems for building and maintaining software, often automatically.

“The power user gap” deserves a closer look. It shows Lorenz curves for AI lines of code, spending and token consumption, with Gini coefficients of 0.77, 0.75 and 0.72. We usually see Lorenz curves and the Gini coefficient in discussions of economic inequality. This is a similar imbalance: a small group accounts for a disproportionate share of AI work, spending and tokens. The top 5% generate roughly 47% of AI lines of code, 46% of spending and 40% of tokens.

It gets more interesting:

• p99 developers produce 46 times as many AI lines per day as the median active user, and 15 times as many merged PRs per week as the median active PR author.
• p90 developers show a smaller gap: 10x for AI lines of code and 4x for merged PRs.

This is not “every developer became 15% faster.” It is a very heavy-tailed distribution, for which an average is not particularly informative; even the median is more useful.

We should not automatically celebrate power users, though. We cannot say that “the best developers became 46 times faster.” These measures mix task types, codebase size, team roles, trust in the tool, context quality, permissions, the habit of splitting work into smaller pieces, the ability to run agents in the background and review culture. Still, the chart is interesting from a management perspective: it suggests what a new operating model could look like.

Someone who can clearly assign tasks to agents, supply context, read diffs, run tests, limit scope and quickly remove blockers may pull far ahead. Not through a secret button, but through a working cycle: intent -> context -> agent run -> verification -> commit/PR -> feedback. For another developer, the same Cursor may remain expensive autocomplete that sometimes gets in the way.

For AI4SDLC, average usage is not enough. We need cohort-level measures to understand:

• Who actually takes AI work through to a PR?
• Where do large diffs appear?
• What happens to review and testing workloads, rework and the change failure rate?
• What distinguishes top users? Do they frame tasks better, know the codebase better, use agents and automations more often, work in repositories with more permissive settings, or simply tackle different kinds of tasks?

If you have power users, talk to them, understand their practices and spread the best ones. But you cannot just look at the tail of the distribution and declare power-user mode the new norm for everyone. That will not work :)

P.S. Our analysis of Cursor’s study is on ai4sdlc-research.space.

#AI #AI4SDLC #Engineering #Agents #DevTools #Management #Metrics