OpenAI tested its AI models on moving COBOL code into languages including Java, Python and Rust while preserving the software’s behaviour. GPT-6.1 Sol scored 87% at low reasoning effort and 98% at higher effort in the company’s evaluations.
COBOL remains embedded in banking and government systems, where replacing old software risks disrupting essential services. Anthropic has also pitched Claude Code for COBOL migration, but provided no performance benchmarks for that work.
For companies modernising legacy software, translating code addresses only part of the task: data, integrations and existing workflows also matter. Experts recommend robust testing and human verification, and caution that OpenAI’s evaluations do not establish that AI can safely replace complex legacy systems.