The takeaway: A GitHub issue documenting a strange pattern in GPT-5.5's reasoning — output clustering at exactly 516 tokens, which reproducers link to wrong answers — hit the Hacker News front page today. The underlying report was filed on 27 June and has since picked up independent reproduction and corroborating issues. OpenAI has not responded. The pattern is well documented; the explanation is not — and that gap is the story.

The pattern

On 27 June, a developer filed an issue against OpenAI's Codex CLI with an unusually large dataset behind it: 390,195 API responses across 865 coding sessions, collected between February and June. The finding: GPT-5.5's reasoning_output_tokens — the hidden "thinking" the model does before answering — cluster at exactly 516 tokens, and at near-multiples of it (1,034, 1,552), far too precisely and too often to be chance.

The numbers in the report are stark. GPT-5.5 accounts for 19.3% of the logged responses but 82% of the exact-516 events. Where other models land on exactly 516 about 1.3% of the time they exceed it, GPT-5.5 does so 44% of the time — roughly 33 times the base rate. And the clustering grew over time: of the runs that reached at least 516 reasoning tokens, the share stopping at exactly 516 rose from 0.11% in February to 53.3% in May (easing to about 36% in the partial June data). The filer explicitly stops short of claiming proof of truncation — the stated claim is the clustering anomaly itself — but suggests it may be behind degraded performance, pointing to a related report in which runs ending at exactly 516 reasoning tokens returned the wrong answer.

Today the issue reached the Hacker News front page, where two commenters reproduced the core observation independently: one ran ten runs of an identical prompt, four of which stopped at exactly 516 reasoning tokens — and all four of those produced a wrong solution; another saw five runs of a logic puzzle all stop at exactly 516, all with incorrect answers. It joins a related issue reporting the same 516-token boundary on failing tasks, earlier degradation complaints, and a report of Codex silently downgrading sessions to a mini model.

What it might be — and what nobody knows

The suspicion running through the thread is that OpenAI is quietly capping reasoning to manage compute costs — truncating the model's thinking budget at a fixed boundary and serving the result anyway. That would square with the wrong-answer correlation: cut a model off mid-reasoning and it answers with whatever it has.

But it's worth being precise about what's established. The pattern is now multiply documented — a large dataset, independent reproduction, several corroborating issues. The explanation is speculation. Hacker News commenters offered benign alternatives: a batching artefact, speculative decoding, a token-budget parameter being set somewhere in the stack, or a misconfigured adaptive-thinking feature. Any of these could produce suspicious round numbers without anyone deciding to throttle anything.

OpenAI has not commented — not on the issue in the eight days since it was filed, and not in the Hacker News thread. Its GPT-5.5 launch page says the model "delivers better results with fewer tokens than GPT-5.4" — which is either the innocent explanation for lower reasoning counts, or an awkward framing for a cap, depending on what the answer turns out to be.

The second silent-throttling story this week

If this feels familiar, it should. On Friday we covered the backlash over Anthropic's silent model downgrades — flags in Claude Code that rerouted requests to cheaper models without telling the user. The pattern repeating across vendors is the point: as inference costs bite, the economics of metered AI create constant pressure to serve less compute than the customer thinks they're getting — and what model actually answers you, and how hard it thinks, is invisible unless someone logs 390,000 responses and checks.

That's also why this report lands differently from the unreproduced Claude Code cache-leak claim we flagged this weekend: this one has data, reproduction, and a specific measurable signature. It still isn't proof of intent. But it's a well-formed question aimed at a company that, eight days in, hasn't answered it.

No mainstream outlet has covered this at the time of writing; sourcing is the GitHub issues and the Hacker News thread, linked above. We'll update if OpenAI responds.