GPT-5.5 Suddenly Got Dumber
Why does GPT 5.5 in Codex stop at fixed token counts like 516 and 1034 on complex tasks? Evidence shows shorter reasoning leads to more wrong answers. Learn the truth and how to protect your work.
“AI Disruption” Publication 10,200 Subscriptions 20% Discount Offer Link.
These past couple of days, a heated issue has blown up in the Codex community.
The core question is simple: Why, even when users explicitly select gpt-5.5 + xhigh, do some complex tasks only think for a brief moment before immediately spitting out an answer?
Even more striking is that many local logs show the exact same number: 516.
On GitHub, someone opened a very technical issue whose title roughly translates to: “gpt-5.5 in Codex exhibits fixed reasoning-length clustering at 516 / 1034 / 1552, potentially linked to degraded performance on complex tasks.”
The issue was still open at the time of writing. The author analyzed Codex local records from February 1 to June 27, 2026 — a total of 390,195 response-level token records across 865 sessions.
The most glaring numbers are:
