GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance (github.com)

Recent findings indicate that the GPT-5.5 Codex model is experiencing degraded performance on complex tasks due to a clustering anomaly in reasoning token outputs. Specifically, there is a disproportionately high incidence of responses hitting exactly 516 reasoning tokens, as well as fixed spikes at 1034 and 1552 tokens. This pattern suggests potential limitations in the model's processing capabilities, especially compared to earlier versions. Notably, while GPT-5.5 accounts for only 19.3% of total responses, it constitutes a staggering 82.0% of instances where responses end at 516 tokens, raising concerns about its reasoning capacity. This discovery is significant for the AI/ML community as it highlights a potential issue with how GPT-5.5 manages reasoning tokens, which could point to underlying problems with its architecture or operational thresholds. The observed decline in overall reasoning-token intensity alongside the spike in token clustering calls for immediate investigation to determine whether this behavior reflects a reasoning-budget enforcement, a truncation issue, or simply an internal threshold. As the model is already being evaluated against various tasks, further scrutiny may clarify the implications for its deployment in high-stakes applications where accurate reasoning is critical.