GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance

324 points · 124 comments on HN · read original →

Points and comments are a snapshot, not live.

GPT-5.5 Codex responses show reasoning token clustering at fixed values, linked to degraded performance on complex tasks.

Analysis of 390k Codex responses shows GPT-5.5 responses disproportionately hit exactly 516 reasoning tokens (82% of such events vs 19% of all responses). Exact-516 rate rose from 0.11% in Feb to 53% in May, while mean reasoning tokens dropped from 268 to 107. User suspects a reasoning-budget cap or truncation, noting related issue #29353 where 516 tokens corresponded to wrong answers.

What commenters are saying

Commenters widely confirm degraded performance with GPT-5.5 Codex xhigh. One user reproduced the issue: 4/10 runs had exactly 516 reasoning tokens, all with wrong answers. Many switched to GPT-5.4 or Claude. Speculation points to OpenAI cost-cutting: a rumor about halving inference cost via thresholded reasoning budgets. Some note the 512 multiple suggests throughput optimization batching. Sceptics suggest this is a business decision to downgrade performance.