OpenAI says its models are improving at COBOL migration for banks

The benchmark results point to faster modernization of bank systems that still run on COBOL, though the company says expert review is still needed.

OpenAI has published results suggesting its models are getting much better at working with legacy languages such as COBOL, which matters because many banks still run critical systems on it and face expensive migrations. The company says around 40% of the world’s banks still depend on COBOL, and it used a new benchmark, MultiLangBench, to test whether models could build and migrate software across old languages. In those tests, GPT‑6 Astra reached an average 87.4% pass rate across 20 languages, while COBOL was still the toughest case at 72%. For migration tasks that preserved program behavior, GPT‑6.1 Sol scored as high as 97.9% with high reasoning effort and 86.6% at low reasoning effort. OpenAI says the results may help teams modernize older systems, but production use still needs expert review and testing against business rules and dependencies.