According to @OpenAI, 30% of SWE-Bench Pro tasks are broken, undermining frontier coding evals and prompting a retraction of its prior recommendation.