Better AI models shift developers toward more complex work
A Cursor-backed study of 500 companies found that improved AI models drove a 44% rise in weekly AI usage and pushed developers toward higher-complexity tasks like architecture, documentation, and code review.
Score breakdown
Developers and engineering teams should expect that adopting more capable AI models will expand — not just accelerate — their workload, particularly in high-overhead areas like architecture, documentation, and code review.
- 01Cursor and Professor Suproteem Sarkar (University of Chicago Booth School of Business) studied developer behavior at 500 companies from July 2025 through March 2026.
- 02The study period included the releases of `Opus 4.5` and `GPT-5.2`, described as step-change advances in AI coding capability.
- 03Average weekly AI messages per user increased 44% over the eight-month study period.
A research paper authored by Luke Melas-Kyriazi at Cursor, conducted in partnership with Professor Suproteem Sarkar from the University of Chicago Booth School of Business, examined how AI model improvements changed developer work patterns across 500 companies using Cursor from July 2025 through March 2026. The study period covered the releases of `Opus 4.5` and `GPT-5.2`, described as delivering step-change advances in AI coding capability. The central finding is consistent with a Jevons-like effect: rather than reducing the need for AI assistance, efficiency improvements drove greater total consumption, with average weekly messages per user increasing 44% over the period.
Developers first applied better models to work of similar difficulty, and only after a 4–6 week lag began tackling more complex tasks.
The complexity shift was not immediate. Developers first applied better models to work of similar difficulty, and only after a 4–6 week lag began tackling more complex tasks. Low-complexity messages grew 22% over the study period, while high-complexity messages grew 68%, with most of that growth concentrated in the final six weeks. The paper attributes the delay to the time developers need to discover new model capabilities and for firms to restructure workflows around them.
Industry adoption was uneven. Media and advertising led with a 54% increase in messages per user, followed by software and developer tools (+47%) and finance and fintech (+45%). The paper hypothesizes that finance may be driven by competitive arms-race dynamics, while media benefits from greenfield opportunities unlocked by more capable models. At the task level, documentation (+62%), architecture (+52%), code review (+51%), and learning (+50%) saw the largest gains — consistent with the idea that as AI-generated code expands codebase size, the overhead of understanding, documenting, and managing that code grows proportionally.
Key facts
- 01Cursor and Professor Suproteem Sarkar (University of Chicago Booth School of Business) studied developer behavior at 500 companies from July 2025 through March 2026.
- 02The study period included the releases of `Opus 4.5` and `GPT-5.2`, described as step-change advances in AI coding capability.
- 03Average weekly AI messages per user increased 44% over the eight-month study period.
- 04High-complexity messages grew 68%, while low-complexity messages grew only 22%; most high-complexity growth occurred in the final six weeks.
- 05Developers showed a 4–6 week lag before shifting from similar-complexity work to more complex tasks after model improvements.