Codex telemetry reveals GPT-5.5 disproportionately terminates responses at exactly 516 reasoning tokens, with additional spikes around 1034 and 1552. This pattern emerged across 390,195 responses analyzed from February through June 2026.
GPT-5.5 accounts for only 19.3% of Codex responses but captures 82% of exact-516 events, with a 44% exact-516 ratio compared to 0.34% for competing models. The clustering intensified from 0.11% in February to over 50% by May, while average reasoning tokens dropped from 268 to 107.
The fixed-value clustering suggests artificial budget constraints rather than complexity-driven reasoning. Mean reasoning tokens fell while exact-516 spikes rose, indicating potential truncation behavior. Other models show negligible clustering at these thresholds, making this GPT-5.5-specific.
**Users should be aware that 516-token responses may indicate artificially constrained reasoning rather than optimal problem-solving, potentially impacting complex task performance.
Source: Hacker News · Summarized by HeadlinesBriefing