kapynResearch

OpenAI finds roughly 30 percent of popular AI coding test is broken

OpenAI found 30% of popular AI coding benchmark SWE-Bench Pro is broken. The company is withdrawing its endorsement, highlighting challenges in accurately evaluating AI coding capabilities. This discovery necessitates a re-evaluation of current benchmarks and a push for more robust testing methodologies.

The Decoder·Jul 9, 2026

Opening Kapyn…