OpenAI Lifts COBOL Code Migration Accuracy to 98%

OpenAI, the company behind ChatGPT, says its new GPT-6.1 Sol model has raised the accuracy of translating COBOL code into modern programming languages. COBOL is a language that is already several decades old, and it still runs inside banking systems today. Accuracy climbed from 5.4% to 86.6% at the same reasoning depth, and up to 97.9% at the maximum effort level. COBOL still powers more than 40% of online banking systems, 80% of offline card payments, and 95% of ATM transactions, and the number of programmers who know the language keeps shrinking every year.

Migrating legacy code is a task AI agents usually handle poorly. Business rules in old systems often exist only inside the code itself. A common mistake is rewriting it with different number rounding, which produces code that runs but calculates money wrong. OpenAI calls this benchmark "close to saturation," but still says migrated code needs a manual review before it goes to production. Even a nearly 98% test score does not remove that review step. But closing this kind of gap on a rare legacy language changes the picture for teams that have spent years afraid to let agents near COBOL.