SWE-bench Is Broken: Why a 90% Score Means Almost Nothing

OpenAI stopped reporting SWE-bench Verified after an audit found 138 flawed tasks, over 60% unsolvable. What broke — and what a real benchmark looks like.