“In a new report, OpenAI audited the SWE-Bench Pro coding benchmark and found that 30% of its tasks were broken.”