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

Article summary

The GPT-5.5 Codex model exhibits a clustering phenomenon where reasoning_output_tokens cluster at fixed values, specifically 516, 1034, and 1552, which may be related to degraded performance on complex tasks. This issue is specific to the GPT-5.5 model and is not observed in other models such as GPT-5.2 and GPT-5.4. The clustering is not evenly distributed across models, with GPT-5.5 accounting for 82% of exact-516 events despite only accounting for 19.3% of responses. The cause of this phenomenon is unknown and may be related to a reasoning-budget, routing, truncation, fallback, or scheduler behavior.

Main themes

• GPT-5.5 Codex performance issues
• Reasoning-token clustering
• Model comparison
• AI performance degradation
• Coding and development tools

What commenters say

• The GPT-5.5 model's clustering phenomenon is likely due to a throughput optimization technique, such as batching reasoning inference in multiples of 512 tokens.
• The issue may be related to a reasoning budget parameter adjustment, which could be a dishonest way of scaling to demand during peak hours.
• Some users have switched to alternative models, such as Claude, due to the perceived degradation in performance of the GPT-5.5 model.
• The use of custom harnesses, such as OpenCode, can provide a workaround for the performance issues, but may not be a long-term solution.
• The problem may not be technical, but rather a business decision to downgrade performance due to cost considerations.
• The degradation of GPT-5.5 reliability may be a deliberate decision to encourage users to upgrade to newer models, such as GPT-5.6.
• Some users have found that using older models, such as GPT-5.4, can provide more reliable performance, while others have turned to alternative solutions, such as running their own models on local hardware.