tl;dr: Analysis of 390K Codex telemetry records shows GPT-5.5 responses disproportionately terminate at exactly 516 reasoning tokens (with additional spikes at 1034 and 1552), accounting for 82% of such events despite being only 19% of traffic. The clustering rose sharply from 0.11% in Feb 2026 to 53% in May 2026, coinciding with a drop in mean reasoning-token usage—suggesting a possible hidden reasoning-budget cap, truncation, or routing threshold that may explain degraded performance on complex tasks.