The Massive Gap in Developer Output
AI coding assistant Cursor recently published an extensive report analyzing two years of aggregated developer usage data. The statistics provide a clear look into how artificial intelligence is actively changing software engineering workflows.
The most striking revelation is the sheer volume of code generated by power users. While the median developer on the platform generates a respectable 700 lines of code per week, the top one percent of users produce between 30,000 and 40,000 lines in the same timeframe. This means elite users are generating the equivalent output of 45 median developers combined. This staggering difference raises immediate questions about code quality. It remains unclear if these power users are building highly valuable greenfield applications, or if they are simply generating massive amounts of boilerplate that will eventually become a maintenance burden.
The Dominance of Input Tokens
Despite the hype around AI writing entire applications, Cursor's data shows that developers primarily use the tool for reading rather than writing. A massive 90 percent of all token usage on the platform consists of input tokens.
This metric aligns perfectly with traditional software engineering principles. Engineers have always spent significantly more time reading existing codebases than typing new characters. Developers are currently feeding large blocks of existing architecture and documentation into the context window, asking the AI to analyze, explain, or refactor the logic.
The Economics of Context Caching
Because input tokens dominate usage, managing their cost is critical. Foundation models charge drastically different rates for reading versus writing. For example, Opus 4.7 charges 5 USD per million input tokens, but the cost spikes to 25 USD per million for output tokens.
To prevent infrastructure costs from spiraling, Cursor implements highly aggressive context caching. By reusing previously processed context, Cursor limits expensive output tokens to just 0.6 percent of total token volume. Context caching accounts for 90 percent of the remaining usage. This efficiency proves that any enterprise building an internal AI agent must implement a robust caching layer, as raw API calls are far too expensive for daily development workflows.
The Rapid Normalization of Automated Commits
The report also highlights a dramatic shift in developer trust. Just a month after highly capable models like Opus 4.7 and GPT 5.5 were introduced, the percentage of developers allowing AI agents to commit code directly without a manual review step jumped from 10 percent to 40 percent.
This metric suggests that as models cross a specific capability threshold, developers are increasingly willing to treat them as autonomous junior engineers rather than simple autocomplete tools. If a model is significantly cheaper (like Composer 2.5) but still produces code that passes automated tests, engineering teams are rapidly adapting their workflows to let the machine take the wheel.