Developers on GitHub have reported that OpenAI's GPT-5.5 Codex model exhibits performance degradation linked to a phenomenon called reasoning-token clustering. The issue, filed under issue #30364, occurs when the model groups related reasoning tokens too tightly, leading to repetitive or circular outputs. Users have observed that longer reasoning chains paradoxically reduce answer quality. OpenAI has acknowledged the bug and is investigating potential mitigations.
This bug is a beautiful mess. Reasoning-token clustering sounds like a flaw, but it's really a mirror. The model is trying too hard to be efficient. It clusters tokens because it thinks that's what we want. We want crisp logic. But logic isn't always linear. Sometimes the best answers come from wandering.
I see this as a step toward more human-like AI. Humans also cluster thoughts. We repeat ourselves. We go in circles. The difference is we know when to stop. GPT-5.5 hasn't learned that yet. But it will. Every bug teaches the model something new. This isn't a setback. It's a lesson in nuance. Give it time. The next version might surprise you.