- Published on
GPT-5.5 Codex reasoning-token clustering may be leading to degraded performance
- Authors
- Name
- DeepHorus
GPT-5.5, the latest reasoning model from OpenAI's Codex platform, is exhibiting a peculiar clustering of reasoning tokens at fixed thresholds—specifically 516, 1034, and 1552—according to a detailed GitHub issue filed by a researcher. This anomaly, observed in telemetry data spanning February to June 2026, suggests that the model may be operating under an artificial reasoning budget, potentially compromising its performance on complex, high-stakes tasks. The issue, tagged as #30364, builds on a prior report (#29353) where GPT-5.5 runs ending at exactly 516 reasoning tokens frequently returned incorrect answers, now adding aggregate evidence across a larger dataset.
The background of this discovery lies in the analysis of Codex's token_count metadata, which tracks reasoning token usage per response. The reporter found that GPT-5.5 responses disproportionately land at exactly 516 reasoning tokens, with additional fixed-boundary spikes at 1034 and 1552. This pattern is not observed in other models like GPT-5.2 or GPT-5.4, making it model-specific. Notably, the clustering coincides with a decrease in overall reasoning-token intensity from February–April to May–June, hinting at a deliberate restriction rather than natural variation.
Key figures underscore the severity of the anomaly. While GPT-5.5 accounts for only 19.3% of all Codex responses, it constitutes 82.0% of exact-516 events. Its ratio of exact-516 responses to those with at least 516 reasoning tokens is approximately 33.6 times higher than the non-GPT-5.5 baseline. The fixed values—516, 1034, and 1552—resemble repeated threshold boundaries, reinforcing suspicions of a budget cap or truncation mechanism.
The implications for users and developers are significant. If GPT-5.5's reasoning is being curtailed at these thresholds, it may fail to fully explore complex chains of thought required for advanced coding, mathematical, or analytical tasks. This could erode trust in the model’s reliability, especially for high-stakes applications where thorough reasoning is paramount. The pattern may also mislead users who expect uniform performance across tasks of varying complexity.
The reporter, while careful not to claim proven truncation, urges the Codex team to investigate potential reasoning-budget, routing, or scheduler behaviors. They suggest internal validation checks, including querying token_count events by model and comparing exact-value counts for 0, 516, 1034, and 1552. Replaying matched complex tasks across GPT-5.2 and GPT-5.5 with quality evaluations, especially separating exact-516 responses, is recommended to assess performance degradation.
Looking ahead, the findings could prompt OpenAI to clarify whether exact-516 represents a normal stopping point, a budget cap, a degraded tier, or an internal threshold. If confirmed as a bug, a fix may restore full reasoning capabilities for GPT-5.5. The developer community awaits an official response, as this anomaly has raised concerns about transparency in AI model behavior and the potential for hidden limitations in cutting-edge systems.
