OpenAI's Codex model exhibits a pattern of reasoning token counts that cluster at exact values, most notably 516 tokens for GPT-5.5, which coincided with lower overall reasoning token usage and may explain degraded performance on complex tasks.

The finding follows issue #29353, which documented a task‑level reproduction where GPT‑5.5 outputs ending at exactly 516 reasoning tokens produced incorrect answers; this analysis aggregates data from February to June 2026 to assess the prevalence of the pattern.

Analysis of 390,195 response records across 865 sessions showed 3,363 instances where reasoning_output_tokens equaled 516; GPT‑5.5 accounted for 19.3 % of all responses but 82.0 % of the exact‑516 cases, yielding a ratio of 44.0 % for GPT‑5.5 versus 1.3 % for other models. Monthly data revealed a sharp rise in the exact‑516 proportion from 0.11 % in February to 53.30 % in May, while mean P90 reasoning tokens fell from 268.1 in February to 106.9 in May.

The author requests that the Codex team examine whether a reasoning‑budget, routing, truncation, or scheduler mechanism causes responses to terminate around 516, 1034, or 1552 tokens, and clarify whether the exact 516 count represents a normal stopping point or an internal threshold.