According to CoinMeta, OneMillion.AI reported that AI, a data company incubated by yc, launched a programming evaluation Real-Swe specifically designed to test the private code of real companies. The tasks were directly sourced from corporate production environments and included tax calculations, bill migration, API billing, and customer data migration. The results showed that Fable ranked first with a pass rate of 38.8%, Astra had a pass rate of 33.8%, Gemini at 38%, and Flash at 31.2%. GLM came in fourth with a pass rate of 28.8%. There were six tasks with an overall pass rate of less than 15%, and one task where all models failed. Zhipu researcher Xiaopu Peng stated that this evaluation further tests Real-Swe's ability to continuously read code, identify patterns, and make multiple changes in unfamiliar projects.