GitHub issue says OpenAI's GPT-5.5 Codex cuts off reasoning at a fixed 516-token ceiling

A GitHub issue filed June 27, 2026 in OpenAI’s own Codex repository reports that GPT-5.5 Codex disproportionately terminates its internal reasoning at exactly 516 tokens. The filer says the ceiling lines up with wrong answers on complex tasks in the coding model.

The report matters because reasoning tokens are the hidden steps a model generates before answering, and a hard ceiling on them would cap how much a coding assistant can work through a problem. The analysis draws on 390,195 response-level token records across 865 sessions logged from February 1 to June 27, 2026, with secondary clustering at 1,034 and 1,552 tokens — each a near-multiple of 516.

The skew is heavily concentrated in one model. GPT-5.5 accounted for 82 percent of all exact-516-token events while making up only 19.3 percent of sampled responses, the issue says — a 44 percent exact-516 rate for GPT-5.5 against 1.3 percent for other models. The clustering rate climbed from 0.11 percent of responses in February to 53.3 percent in May, while the mean reasoning-token count fell from about 268 to 107. A companion issue reported that GPT-5.5 runs ending at exactly 516 tokens returned wrong answers.

The caveats are large. The filing is an unverified developer bug report, not an OpenAI-authored finding, and the author says plainly the data does not prove hidden chain-of-thought truncation and should be read as a telemetry anomaly, not a confirmed defect. OpenAI has not responded to the issue as of this writing, and the pattern has not been independently reproduced.

Whether the clustering reflects a genuine regression, a logging artifact or an intentional token budget remains open until OpenAI weighs in.