cat posts/2026-10-29-stack-overflow-under-a-thousand.md

Stack Overflow, under a thousand

· 4 min read · ai community stack-overflow knowledge

When “StackOverflow now gets fewer than 1000 questions a month” went up on Hacker News last night, the first thing I did was check which month it was talking about. In January the viral number was 300 a month, and it came from someone charting a month that was five days old, so I went in expecting the same mistake.

It’s not the same mistake. The link goes to a Data Explorer query, which I couldn’t get to load (Cloudflare kept asking whether I was human, which felt a bit pointed given the topic), so I counted with the public Stack Exchange API instead, which will give you a total of questions for any date range. September, all thirty days of it, had 969 new questions. August had 1,019. October, through yesterday, has 918, which puts it on course to finish within a few dozen of a thousand either way. Data Explorer only refreshes about once a week, so if the last point on that chart is October it’s a partial month again, but this time it doesn’t matter, because September gets under the line on its own.

1,000 3,615 3,043 2,987 2,322 1,986 1,701 1,175 1,019 969 918 Jan Feb Mar Apr May Jun Jul Aug Sep Oct* New questions per month in 2026, Stack Exchange API (deleted questions aren't counted)
*October is the 1st through the 28th. The January post's chart had March 2014 at about 207,000 on its own; on that scale the tallest of these bars would be about three pixels high.

The shape is different from what I expected. Through July it lost a few hundred questions most months, more than 500 in three of them, and since August it has barely moved. August, September and most of October all sit around a thousand, which looks like a floor to me, though I’d have said the same about June.

To put a thousand in some kind of scale: at the March 2014 peak someone asked a question about every 13 seconds, and in September it was about one every 45 minutes. One commenter on the thread pointed out that it’s now fewer than the site’s first month, and that checks out too. August 2008, when Stack Overflow was still an invite-only beta, had 3,744.

Most of the thread was about moderation, closed-as-duplicate stories and people who gave up asking years ago. One commenter was “confident that its horrid moderation drove people away long before AI had a chance to.” I agree with half of that, and I said in January that the slide started around 2014. But a moderation problem that’s been there for twelve years doesn’t explain losing nearly three quarters of what was left between January and September.

I wanted to know who’s still there, so I pulled all 969 September questions. Python has the most (113) and C++ is second (84). The highest-scored ones are mostly the kind of question you want a person for, things like what constexpr atomic<T>::wait() does at compile time, or how C++26 hives keep iterators valid when elements are erased. Language-lawyer questions, basically, where the answer depends on reading the standard carefully and a model will happily tell you something plausible and wrong. Another 92 are tagged advice, best-practices or tooling-recommendation, which is the sort of question the site spent fifteen years closing as opinion-based. And a bit more than a fifth were asked by people with over 1,000 reputation, so a good chunk of who’s left are old regulars.

The answering side looks healthier than I thought it would. About three quarters of September’s questions have at least one answer, and about a third have an accepted one. The hives question was answered less than an hour and a half after it was posted, by someone with over 120,000 reputation, and the answer was accepted.

My own team’s questions don’t go anywhere near the site. Questions about our code go to Claude Code in the repo, and everything else goes to the internal chat page we’ve had since 2023. I can’t remember the last time anyone pasted a Stack Overflow link in a PR, and it used to happen weekly.

In January I said my small fix was to write down anything an agent and I worked out that wasn’t in the docs, somewhere other people could find it. I’ve kept that rule maybe a third of the time. A few issue comments on open source projects, a handful of pages in our wiki, and a lot of sessions where I fixed the thing, felt pleased and closed the terminal. The agent would have drafted the write-up in a minute. I just didn’t ask it.

Someone in the thread asked whether model answers will get worse without public sources like this. I don’t know. The hives question has been viewed about 1,700 times in seven weeks, which is more readers than any of my agent sessions will ever have.