GPT-5.5 Codex Shows Anomalous Clustering in Reasoning Tokens

petrusenko_max · x · 2026-07-05

Analysis reveals that GPT-5.5 Codex outputs show anomalous clustering at 516, 1034, and 1552 reasoning tokens, a model-specific pattern linked to reduced reasoning intensity. This implicit token cap is likely the exact reason why the model struggles with complex tasks.

Related event: GPT-5.5 Reportedly Degrades Due to Silent Token Truncation(3 posts)→

Original post →

More from Models

Models channel →