GPT-5.5 Codex keeps stopping at exactly 516 reasoning tokens and getting it wrong
Give GPT-5.5 Codex a puzzle and sometimes it thinks for exactly 516 tokens, then hands back the wrong answer. An OpenAI Codex issue shows those stops aren't random: they pile up at fixed thresholds 518 tokens apart, and they track failures.
Across 390,195 logged responses, GPT-5.5 Codex terminates reasoning at exactly 516 tokens far more than any sibling: 82% of exact-516 events are GPT-5.5, though it's only 19% of traffic. The stops cluster at 516, 1034, and 1552, each about 518 apart.
One HN user reproduced it live: 4 of 10 identical prompts short-circuited at 516 tokens, and every one returned the wrong answer.
The suspected cause: a throughput optimization batching reasoning in multiples of 512.
Reasoning that stops at exactly 516 tokens
Someone filing against OpenAI's open-source Codex repo pulled 390,195 response records from 865 sessions and found GPT-5.5's reasoning does not taper naturally, it slams into walls. Responses terminate at exactly 516 reasoning tokens far more than chance allows, with secondary pileups near 1034 and 1552, each roughly 518 apart. GPT-5.5 accounts for 82% of exact-516 events while making up only 19.3% of all responses. Over time the clustering exploded: from 0.11% of responses in February to 53.30% in May, even as mean reasoning tokens fell from 268 to 107. The stuck responses correlate with wrong answers on hard tasks.
The pattern is specific to this one model.
| Model | Exact-516 / ≥516 ratio |
|---|---|
| GPT-5.5 | 44.0% |
| GPT-5.4 | 19.8% |
| GPT-5.2 | 0.34% |
| GPT-5.3-codex | 0.0% |
No OpenAI reply is on the issue yet.
Four in ten runs short-circuit
In the thread, a reader turned the statistics into a live reproduction, feeding one puzzle prompt through the Codex CLI ten times.
"Out of 10 runs of the exact same prompt, 4/10 had this 516 thinking token issue, and every one of these had the wrong solution. So nearly half the time, 5.5 xhigh could be short circuiting and degrading performance." — a user who reran the identical prompt ten times
The most-repeated explanation reads it as an infrastructure choice, not a model change.
"Clearly they are batching reasoning inference in a few multiples of 512 tokens as a throughput optimization." — a commenter reading it as a batching optimization
That lands against a memory another reader raised: a recent claim that OpenAI had halved its inference compute with an optimization breakthrough. If the budget cap is the savings, users are paying in silent quality drops. One heavy xhigh user described exactly that.
"I'm seeing incredibly stupid implementations intermittently, and have simply switched to Claude until openai takes the issue seriously." — a daily xhigh user who left for Claude
The reproduction cuts one way that matters: because Codex is open source, an outside user could log token counts, run the tests, and post the histogram in public, a diagnosis OpenAI never volunteered. The same silent server-side change on a closed model would be invisible.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Anna's Archive offers $200k to exfiltrate every Google Books scan
A pirate library has put a price on the largest book-scanning project ever built: $200,000 for a full copy of Google Books' scans, paid to whoever delivers them.