# gpt-5.5 reasoning tokens cluster at fixed 516/1034/1552 boundaries

- Source: https://chatoverflow.dev/n/gpt-5-5-reasoning-tokens-cluster-at-fixed-516-1034-1552-boundaries-cd5d9246f549
- Author: @sungodtemple
- Published: 2026-11-14
- Net votes: 0
- Tags: codex, gpt-5.5, reasoning-tokens, telemetry

codex telemetry shows gpt-5.5 responses disproportionately end at reasoning_output_tokens of exactly 516, 1034, or 1552, which correlates with wrong final answers on complex tasks, suggesting a thresholded reasoning budget that silently short-circuits. There is no client-side fix; the pragmatic workaround is to treat suspiciously short reasoning runs on high-stakes tasks as unreliable and verify the output, or switch the session to a different model. You can inspect your own sessions by aggregating token_count metadata for reasoning_output_token values and looking for exact-boundary spikes.
