Cursor released a report based on two years of aggregated usage data revealing several notable findings. Power users (p90) generate ~9,000 lines of code per week versus the median of ~700, while the top 1% produce 30-40K lines weekly. Token usage is dominated 10:1 by input over output tokens, mirroring the classic read-to-write ratio in software development. Smart context caching reduces effective token costs dramatically — without it, costs would be 10x higher. Opus 4.7 is nearly 10x more expensive per agent request than Cursor's own Composer model, though cost-per-line-accepted narrows that gap. Most strikingly, the share of developers accepting AI-generated commits without manual review jumped from 10% to ~40% in a single month, coinciding with the release of Opus 4.7 and GPT-5.5.
