GPT-5.5 shows clustering anomalies affecting Codex performance

Overview of the Issue

A recent analysis highlights a potential performance issue in the GPT-5.5 Codex model. It points to an unusual clustering of reasoning tokens—specifically at 516, 1034, and 1552 tokens—which may lead to reduced effectiveness on complex tasks.

Evidence of Clustering

The analysis indicates that responses landing at exactly 516 reasoning output tokens have significantly increased, while the overall intensity of reasoning tokens has decreased. Notably, GPT-5.5 accounts for a disproportionate amount of these specific responses, suggesting a possible model-specific problem.

Comparative Analysis

Other models do not exhibit the same clustering behavior, indicating that this issue is unique to GPT-5.5. The clustering at fixed values raises concerns about whether underlying processes such as reasoning-budget or fallbacks within the model are functioning correctly.

Call for Investigation

Given these findings, it is recommended that the Codex team investigate potential causes linked to reasoning-budget constraints, truncation, or anomalous scheduler behavior in GPT-5.5. This could help clarify the root of the observed performance degradation.