StackOverflow is down to <1,000 questions a month. The Q&A era is over.

What happened

A query against StackOverflow's own public data explorer, posted to Hacker News on October 29 and climbing to 226 points, shows new question volume has fallen below 1,000 per month for the first time since the site's first full year of operation. In 2014, StackOverflow was taking in roughly 200,000 new questions a month. In October 2026, it's taking in fewer than 1,000. That's a 99.5% collapse in a little over a decade, and the curve isn't flattening — it's still bending down.

The shape of the decline is the interesting part. Volume peaked around 2014, drifted gently downward through the late 2010s, then fell off a cliff in late 2022. The timestamp is not subtle: ChatGPT launched on November 30, 2022. Within six months, question volume had halved. Within two years, it had fallen by an order of magnitude. The 2024 lay-off wave at StackOverflow — roughly 28% of staff — now reads less like a cost-cutting measure and more like triage.

The numbers themselves come from `data.stackexchange.com`, StackOverflow's own query interface. This isn't a leak or a competitor's analysis. It's the site publishing its own vitals, and the vitals are flatlining.

Why it matters

The first-order story is obvious: developers stopped asking StackOverflow questions because they started asking ChatGPT, Claude, and Copilot instead. The LLM you're using right now was trained in large part on the exact corpus whose production it is now suppressing. There's a term for this in ecology — a trophic cascade — and the Q&A ecosystem is in the middle of one.

But the pre-ChatGPT decline matters more than people remember. StackOverflow was already losing ground before GPT-3.5 existed. The reasons were cultural and structural: aggressive moderation that drove off new askers, a reputation system that rewarded gatekeeping over helpfulness, a long-tail problem where "your question is a duplicate of this 2011 thread" became the dominant response. The platform optimized for a canonical knowledge base, which is exactly the wrong optimization when the questions you need answered are novel. LLMs didn't kill StackOverflow. They finished off a patient that had been bleeding out for five years.

The second-order problem is more interesting, and it's the one nobody wants to say out loud. The training data for every major coding model — GPT-4, Claude, Gemini, Llama, DeepSeek — leans heavily on StackOverflow's historical corpus. That corpus was built by humans asking novel questions and other humans answering them in public. If askers stop asking and answerers stop answering, the next generation of models has nothing new to train on for problems that didn't exist in 2022.

Ask Claude about a React 18 Suspense edge case and the answer is strong, because the corpus is thick. Ask about a React 19 `use()` hook interaction with a specific error boundary pattern introduced in 2025 and the answers get vaguer. The models are inheriting a knowledge base that is slowly aging out from under them, and the pipeline that used to refill it is empty. GitHub Issues and Discord picked up some of the slack, but Discord isn't crawlable and GitHub Issues are repo-scoped, not canonical. Neither replaces the public, searchable, labeled Q&A graph.

Community reaction on Hacker News captured the ambivalence well. One commenter: "I haven't asked a StackOverflow question in three years. I also haven't learned anything hard the way I used to." Another: "The site deserved to die for how it treated new users, but what replaces it is strictly worse for the long term." These are not incompatible views. StackOverflow was both a hostile moderation culture and the single most important public knowledge graph software has ever had. Both can be true.

What this means for your stack

First, treat StackOverflow's archive as a frozen asset, not a living reference. The site isn't going anywhere — Prosus, the parent company, won't let it — but it is effectively a historical dataset now. If your team has an internal wiki that links to SO threads, those links are safe. The threads they point to will still be there. The threads you'd want for problems you'll hit next year are the ones that won't exist.

Second, your LLM pair-programmer is going to get subtly worse at new frameworks and newer versions of old ones, and you won't see it coming because the failure mode is confident hallucination, not refusal. The gap between "the model is great at this" and "the model is making things up about this" is now measured in how recently the library landed, and SO's collapse widens that gap every month. For bleeding-edge work, read the actual source, read the actual issues, and don't trust confident model output on APIs younger than the training cutoff.

Third, if you maintain an open-source library, your GitHub Issues are now load-bearing in a way they weren't before. They're where the Q&A that used to go to SO is landing — unstructured, unsearchable across repos, and increasingly the only public record of how your software actually gets used. Consider enabling GitHub Discussions, writing out resolved issues as blog posts, and generally treating issue triage as a form of documentation. Someone's model is going to be trained on this.

Fourth — and this is the uncomfortable one — if you teach junior developers, the muscle of articulating a technical problem clearly enough for a stranger to answer it is being lost. Writing a good StackOverflow question forced you to isolate the minimal repro, state what you'd tried, name what you expected versus what happened. That skill transfers to debugging, to pull requests, to incident reports. Asking an LLM doesn't require it, because the LLM will cheerfully pattern-match on a vague prompt. Juniors who grew up asking Claude never learned to ask the question that unlocks the answer.

Looking ahead

The honest forecast is that StackOverflow in its current form never recovers. The question isn't whether the site rebounds — it won't — but whether anything replaces its function as a public, searchable, canonicalized developer knowledge graph. GitHub Discussions is the obvious candidate but has no cross-repo discoverability. Discord is a dark forest. Reddit is Reddit. The LLM providers have an obvious incentive to build the replacement themselves — a public, model-generated, continuously updated knowledge base — but that's a snake eating its tail, and the output quality curves show it. The next five years of developer tooling will be shaped by what, if anything, fills this hole. Right now, nothing is.