Study: Better AI Models Drive 44% Surge in Developer Usage, Shift Focus to Complex Tasks
Key Takeaways
- ▸AI usage among developers increased 44% after releases of more capable models, supporting the Jevons paradox that efficiency gains drive increased consumption rather than reduced work
- ▸Developers initially performed more existing tasks with better models before shifting to complex work after 4-6 weeks, suggesting a learning and workflow adaptation period
- ▸Finance, media, and advertising sectors led adoption (45%, 54%, and 54% increases respectively), driven by competitive dynamics and new market opportunities
Summary
A new research study conducted by Anthropic in partnership with the University of Chicago Booth School of Business reveals that improvements in AI models significantly increase developer demand rather than merely automating existing work. Analyzing data from 500 companies using Cursor between July 2025 and March 2026—a period that included releases of Opus 4.5 and GPT-5.2—researchers found that average weekly AI messages per user increased 44% during the study period, consistent with a Jevons-like effect where efficiency gains drive increased consumption.
The research identified distinct patterns in how developers adapted to more capable AI models. Initially, developers used better models to accomplish more work of similar complexity, but after a 4-6 week lag, they began tackling significantly more complex tasks. Low-complexity message usage grew 22% while high-complexity usage surged 68%, primarily in the final six weeks of the study period. The shift was most pronounced in finance, media, and advertising sectors, where competitive pressures and new opportunities accelerated adoption.
Crucially, the study found that as AI improves at code generation, developer responsibilities shift away from writing code toward managing AI output. Documentation requests increased 62%, architecture tasks grew 52%, code review climbed 51%, and learning queries rose 50%, while self-contained tasks like UI/styling increased only 15%. This pattern suggests that more capable AI models expand codebase size and complexity, creating new downstream work in documentation, review, and system integration rather than simply replacing developer effort.
- Better AI models shifted developer focus from code generation to higher-value tasks like documentation (+62%), architecture (+52%), and code review (+51%), creating new work rather than eliminating it
Editorial Opinion
This research provides compelling empirical evidence that AI augmentation creates rather than destroys developer productivity and economic opportunity. The finding that more capable models shift developers toward complex architectural and management work—rather than making them obsolete—suggests a promising future where AI handles routine coding while humans focus on strategic decision-making. The industry-specific patterns also highlight how competitive dynamics and market structure influence AI adoption, underscoring that AI's real-world impact depends heavily on business context.
