What happened
OpenAI's GPT-5.5 model within Codex exhibits anomalous reasoning token clustering at fixed boundaries of 516, 1034, and 1552 tokens, according to a GitHub issue. Analysis of 390,195 response records from February to June 2026 revealed GPT-5.5 accounted for 82.0% of exact-516 events despite representing only 19.3% of all responses. This clustering, which sharply increased from 0.11% in February to 53.30% in May, coincides with a decrease in overall reasoning-token intensity, potentially degrading performance on complex tasks.
Why it matters
This token clustering in GPT-5.5 could introduce unpredictable performance ceilings for developers building on Codex, particularly for complex tasks requiring extensive reasoning. Procurement teams and architects must account for potential reasoning budget limitations, as the model's behaviour suggests an internal thresholding mechanism. This impacts reliability and could increase development cycles for robust AI applications, echoing earlier concerns about LLMs failing core cognitive attention tests.
