GPT-5.5 Codex May Be Silently Truncating Its Own Thinking

Telemetry analysis reveals GPT-5.5 Codex reasoning tokens cluster at fixed boundaries of 516, 1034, and 1552, suggesting deliberate reasoning caps that coincide with degraded performance on complex tasks.

A systematic analysis of Codex telemetry data spanning February to June has uncovered a suspicious pattern in GPT-5.5 responses: reasoning output tokens disproportionately land at exactly 516, with additional clustering at 1034 and 1552.

The pattern is model-specific. Earlier Codex models do not exhibit these fixed-boundary spikes. The clustering coincides with lower overall reasoning-token intensity, which the researcher believes may explain degraded performance on complex and high-stakes tasks.

A prior report documented a case where GPT-5.5 runs that ended at exactly 516 reasoning tokens returned wrong answers. The new aggregate analysis strengthens that signal with statistical evidence across a broader time window.

The bug report stops short of claiming hidden chain-of-thought truncation is proven. But the data narrows the range of plausible explanations: thresholded reasoning-budget behavior, checkpointed token allocation, or an internal truncation layer — none of which OpenAI has publicly documented.

With over 240 points on Hacker News, the thread reflects a community increasingly suspicious that model providers are silently capping reasoning depth to manage inference costs — and that users may be paying for thinking their models never actually did.

Sources: GitHub Issue #30364