GitHub issue claims GPT-5.5 Codex shows fixed reasoning-token spikes that may hurt complex task performance

A new Codex GitHub issue reports that GPT-5.5 responses appear to cluster at exact reasoning token counts such as 516, 1034, and 1552. The reporter says the pattern is model-specific, based on Codex token_count metadata from February to June 2026, and may correlate with degraded results on complex tasks, while stopping short of claiming proof of hidden chain-of-thought truncation.

If confirmed, the anomaly would suggest that reasoning budgets or internal thresholds can create measurable failure modes in coding agents. The report also highlights a growing need for better telemetry and debugging tools as developers rely on AI systems for high-stakes programming work.