OpenAI analysis flags flaws in SWE-Bench Pro benchmark methodology and known limitations, exposing accuracy gaps in AI coding agent evaluation benchmarks reliability issues.