GPT-5.5 Codex reasoning token clustering correlates with degraded output

Category: AI
Status: developing
Sources: codex issue 30364, HN discussion
Summary: A community analysis of Codex logs reports that GPT-5.5 Codex reasoning token counts cluster at fixed values (516, then 1034 and 1552, spaced about 518 apart) far more than other models. GPT-5.5 shows an exact-516 rate of 44.0% of runs at or above 516 tokens versus 1.3% for non-GPT-5.5 models, and the exact-516 share rose from 0.11% in February 2026 to 53.30% in May 2026. Runs that stop at exactly 516 reasoning tokens correlate with wrong answers on complex tasks. OpenAI has not responded on the issue.
Comments: HN commenters read the 512-multiple spacing as reasoning inference batching for throughput, and several report intermittent quality drops in Codex over recent months, with some switching models.
Why it matters: If a serving-side reasoning budget truncation is degrading a widely used coding model, it changes how much to trust its output on hard tasks.
Follow-up: Watch for an OpenAI acknowledgment or a serving-side fix.