Cursor, the AI-powered code editor, has released its Spring 2026 Developer Habits Report based on two years of aggregated usage data[reference:18]. The findings paint a picture of how AI is fundamentally changing the way developers write code.
The 1% vs. The Rest
The median Cursor user generates about 700 lines of code per week with AI assistance. The 90th percentile comes in closer to 9,000 lines. But the top 1% of users generate around 30,000 to 40,000 lines of code per week[reference:19]. That's the equivalent of what about 45 median developers generate in the same period[reference:20].
These power users also merge 15 times more pull requests per week than the median user, meaning the code isn't just being generated, it's being accepted[reference:21]. The report raises interesting questions: Are these users writing more greenfield code? Do they generate 45 times as many bugs? Are they adding real business value with the software they ship?
The 10:1 Read-to-Write Ratio
Perhaps the most surprising finding is that 90% of Cursor's token usage is input tokens. Most tokens are used for reading the existing codebase and documentation, not for generating new code[reference:22]. This mirrors the classic 10:1 read-to-write ratio that Robert C. Martin identified in his book Clean Code back in 2008.
When taking cache usage into account, Cursor spends only 0.6% of tokens on output tokens. The remaining 99% is split between cache reads (90%), cache writes (2.5%), and input tokens (7%)[reference:23]. This suggests that context reuse and caching will be key efficiency components for AI coding tools going forward.
Model Economics
At the time of publishing, Opus 4.7 was considered the most capable coding model, but it's close to 10 times more expensive than Cursor's own Composer 2.5 model[reference:24]. Opus 4.7 is twice as expensive as GPT-5.5 and nearly 10 times more than Composer 2.5. However, when measured by cost per line of code accepted, Opus 4.7 performs similarly to GPT-5.5.
Interestingly, Google's Gemini models were left out of the report simply because Cursor sees very little usage of them on their platform, similar to the sparsely used Grok model[reference:25].
The Trust Shift
In just one month, the percentage of developers letting AI agents create commits without a manual step jumped from 10% to around 40%[reference:26]. The jump correlates with the release of Opus 4.7 and GPT-5.5, suggesting that developers are gaining trust in AI-generated code[reference:27]. Large pull requests (defined as at least 1,000 lines of changes) are becoming more common as developers use AI to accomplish larger amounts of work in a single PR[reference:28].
The report concludes that AI coding is moving from a tool-based productivity gain to a systemic restructuring of how software is built[reference:29]. The question isn't whether AI will change development, but how quickly and in what ways.