A statistical anomaly in OpenAI's GPT-5.5 model has been documented in exhaustive detail on the Codex issue tracker, and the numbers are hard to dismiss as coincidence.

A user analyzed 390,195 Codex token-count records across 865 sessions spanning February through June 2026. What they found was a sharp, model-specific clustering pattern: GPT-5.5 responses disproportionately terminate reasoning at exactly 516 output tokens.

The numbers tell the story. GPT-5.5 accounts for only 19.3% of all Codex responses — but 82% of all exact-516 events. When GPT-5.5 produces a response of 516 or more reasoning tokens, 44% of the time it lands at exactly 516. The same ratio for all other models combined is just 1.3%.

The clustering also follows clear threshold boundaries. Spikes appear not only at 516 but also at 1034 and 1552 — values that look like repeated budget caps rather than natural stopping points.

The timing is particularly telling. In February, the exact-516 ratio was 0.11%. By May it had jumped to 53.3%, moderating to 35.8% in June. Over the same period, mean reasoning-token intensity cratered from 268 to 107, and the 90th percentile fell from 772 to 344.

The issue's author stops short of claiming hidden chain-of-thought truncation, framing it as a narrower observation: Codex telemetry shows a GPT-5.5-specific fixed-token clustering anomaly consistent with thresholded reasoning-budget behavior. The ask is straightforward — investigate whether a reasoning budget, routing rule, scheduler, or fallback is causing responses to terminate at these specific boundaries.

If confirmed, the implications extend beyond one GitHub issue. Degraded reasoning depth under the hood, invisible to end users, would mean that some Codex users are getting systematically shallower analysis on complex tasks without knowing it. The thread, still open with 66 comments, has drawn attention on Hacker News where it sits near the top of the front page.

Sources: GitHub Issue #30364, Hacker News Discussion