GPT-5.5 Codex is quietly getting worse

A detailed thread on Hacker News documents a specific and reproducible problem with GPT-5.5 Codex: reasoning output tokens cluster at fixed values spaced exactly 518 apart. The community's read is that OpenAI is batching reasoning inference in multiples of 512 tokens as a throughput optimization, which caps and distorts the model's actual reasoning. Multiple commenters who use the model heavily say they have already stopped using it for reasoning tasks because it simply does not work the way it used to.

The pattern here is a familiar one in platform businesses: a provider makes a backend optimization to cut costs, the change degrades user experience in subtle ways, and users only figure out what happened by reverse-engineering behavior. The fact that someone was able to detect this through token clustering analysis suggests the degradation is structural, not random.

Commenters note this is almost certainly a deliberate throughput tradeoff, not a bug. That framing is actually more alarming for builders: if OpenAI is willing to silently cap reasoning depth to save money, there is no guarantee the model you benchmarked last month is the model running in production today.