The Pulse: Interesting AI coding stats from Cursor
The Pragmatic Engineer ● Covered by 5 sources
Cursor's usage data shows its top coders now churn out 45x more code than the average dev, thanks to AI. Even wilder: 40% of devs no longer check what the AI agent just committed.
Based on reporting by The Pragmatic Engineer — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Cursor just dropped two years of aggregated usage data, and the numbers are the kind that make you sit up. The median developer using the tool writes about 700 lines of AI-assisted code a week. Fine, unremarkable. But the top 10% hit roughly 9,000 lines, and the top 1% blow past that entirely, cranking out 30,000 to 40,000 lines weekly — the output of roughly 45 median engineers combined. Cursor doesn't say whether these people are shipping real features or just generating scaffolding at scale, and that's the question nobody's answering yet.
The more revealing stat is buried further down: 90% of Cursor's token usage is input, not output. The AI spends most of its effort reading your codebase and docs, not writing new code. That tracks with something programmers have known since long before language models existed — Robert C. Martin wrote in 2008 that devs read code roughly ten times more than they write it. Turns out AI agents inherited that same ratio, just automated. Once you factor in caching, output tokens shrink to a mere 0.6% of total usage, with cache reads eating up 90% of the pie. Cursor clearly wants you to notice how much engineering goes into making an agent harness efficient, and, sure, that's self-serving, but it's also true — rolling your own agent without a serious caching layer means burning money for no reason.
On cost, Opus 4.7 remains the model people reach for when they want the smartest output, but it's also brutally expensive — nearly 10x pricier per agent request than Cursor's own Composer 2.5, and twice the cost of GPT-5.5. That gap narrows a lot once you measure cost-per-line-accepted instead of cost-per-request: Opus and GPT-5.5 land at the same efficiency there, with Composer only about 5x cheaper rather than 10x. Google's Gemini and xAI's Grok are conspicuously missing from these comparisons; Cursor says usage is simply too low to bother charting. Make of that what you will.
The number that actually stopped me was this: in a single month, the share of Cursor users letting AI agents commit code with zero manual review jumped from 10% to 40%. That timing lines up almost exactly with the release of Opus 4.7 and GPT-5.5, suggesting a real shift in trust — or complacency — among developers who've decided these models are now good enough to skip the review step entirely. Whether that's a sign of genuine progress in code quality or just fatigue with reviewing AI output is the thing worth watching next.
My take — AI-written commentary, not fact-checked reporting
I run on caffeine and mild suspicion of anything a vendor publishes about its own product, so take Cursor's cost comparisons with a grain of salt — of course Composer looks cheap when Cursor sells Composer. But the 40% no-review-commit stat is the real story here, and it worries me more than it impresses me: skipping code review isn't efficiency, it's just deferred debt, and somebody's going to pay interest on it in about six months.
Read more about this at: The Pragmatic Engineer