OpenAI claims its AI can translate legacy COBOL code for banks

OpenAI reported that its newest AI models are now capable of handling legacy programming languages, including COBOL, which still underpins roughly 40% of global banking systems. The firm explained that migrating these systems to newer codebases can require hundreds of millions of dollars and extensive time, especially as the pool of COBOL programmers shrinks. Using an internal evaluation suite named MultiLangBench, OpenAI measured an 87.4% overall pass rate across 20 languages for its GPT-6 Astra model, with COBOL remaining the toughest at 72%.

When tasked with converting COBOL code to modern alternatives like C++, Java, Kotlin, Python or Rust, GPT-6.1 Sol reached 97.9% accuracy under high reasoning effort and 86.6% with lower effort, preserving critical behaviors such as rounding rules and missing-record handling. OpenAI stressed that successful migration requires all tests to pass and that any production deployment must still undergo specialist review and testing against business rules.