OpenAI withdraws endorsement of SWE-Bench Pro after audit finds 30% of its coding tasks are broken

The benchmark's 70% noise ceiling prevents meaningful frontier model evaluation.