Better AI Models Enable Bolder Work

Tracking the work habits of developers at 500 companies using Cursor from July 2025 to March 2026, improvements in model performance didn't just create simple efficiency gains—they generated more demand. This period saw stepwise performance improvements like Opus 4.5 and GPT-5.2, and average weekly messages increased by 44%.

Initially, developers used the better models to handle more of the same kind of work. But after a lag of 4-6 weeks, a clear shift toward more complex tasks emerged, with low complexity messages increasing by 22% and high complexity messages increasing by 68%.

By industry, media and advertising saw the largest increase at 54%, followed by software and developer tools at 47% and finance and fintech at 45%. The authors attribute this difference to pressure to secure competitive advantage in finance, versus the expansion of new opportunity areas in media and advertising.

The types of work also changed. Tasks dealing with what comes after code generation—such as documentation (+62%), architecture (+52%), code review (+51%), and learning (+50%)—increased significantly, while relatively self-contained tasks like UI/styling (+15%) increased less. As AI generates more code, the developer's role is shifting toward documenting, understanding, reviewing, and coordinating that output.

Ultimately, this research shows that AI can go beyond assisting with existing work to enable larger tasks that were previously out of reach, potentially expanding economic activity itself. The efficiency gains manifested as a Jevons-like effect, where increased efficiency led not to reduced consumption but to expanded total demand.