OpenAI GPT-5.5: an evaluation

OpenAI's GPT-5.5 model demonstrates notable advancements in security, showcasing one of the cleanest vulnerability profiles with a low density of 75 vulnerabilities per million lines of code (mLOC) and a consistent distribution across severity levels. Despite these improvements, the model presents challenges in concurrency, exhibiting a high rate of threading bugs at 170 per mLOC, which are difficult to detect during code reviews and require thorough testing and analysis. The code generated by GPT-5.5 spans over 700,000 lines with minimal comments, resulting in elevated cognitive complexity that increases the burden of human verification. While the model excels in generating secure code, its volume and complexity necessitate a structured approach to code review, especially for concurrency issues, to manage the verification debt effectively.