OpenAI AI Models Boost Legacy Programming Migration With 97.9% COBOL Accuracy
OpenAI released research demonstrating that enhanced models like GPT-6 Astra and GPT-6.1 Sol dramatically improve performance on low-resource legacy programming languages. In evaluations, the models achieved up to a 97.9% accuracy rate when migrating COBOL code to modern languages. This breakthrough addresses a critical enterprise and government vulnerability by making the modernization of aging, core infrastructure significantly faster and more reliable. Automated legacy migration reduces the high costs and severe risks associated with a dwindling pool of developers who understand older systems. Under high reasoning intensity, GPT-6.1 Sol achieved 97.9% migration accuracy for COBOL, while GPT-6 Astra reached an 87.4% average pass rate across 20 languages in MultiLangBench. OpenAI also introduced lower-cost variants like GPT-6 Luna, which reduces estimated API costs by roughly 78% compared to previous generations.
## BACKGROUND
COBOL (Common Business-Oriented Language) and Fortran are older, compiled programming languages created in the late 1950s that still heavily power critical global sectors like finance, banking, and weather forecasting. Because these languages lack large open ecosystems or modern documentation, they are considered low-resource languages that pose major challenges for traditional AI models.