OpenAI says its new models can finally handle COBOL

OpenAI says its newest models can now write and move code in COBOL, Fortran and Visual Basic, languages that most AI coding tools have fumbled until now. If you bank, check a forecast or work anywhere with decades-old software, that code sits closer to your life than you might think. The results come from OpenAI's own tests, so treat them as company claims for now.

OpenAI calls these "low-resource" languages, meaning there is far less code and documentation online for a model to learn from than there is for Python. The company says COBOL handles 95% of ATM transactions, while Fortran turns up in weather forecasting and nuclear reactor safety analysis. Developers and enterprise customers told OpenAI the code is hard to maintain because the original authors are often gone and key business rules live only in the code itself.

One test asks a model to build a SQLite-compatible database from scratch in 20 languages. GPT-6 Astra passed about 87% of the checks, up from roughly a quarter for GPT-5.6 Sol. Those percentages count checks passed, so they show how much of the expected database behavior worked, not how fast, tidy or safe the code is to run. COBOL stayed the hardest language, though it climbed to about 72%. On a separate test that asks a model to move COBOL programs into another language while keeping their behavior, GPT-6.1 Sol came close to a perfect score at higher reasoning effort, which OpenAI says nearly saturates the test.

Both benchmarks are OpenAI's own, and the post offers no outside validation. Passing a database test suite is a long way from safely replacing a live payments system, and the post does not say how or when these models will reach customers in products.

Watch for independent teams to rerun the database test, and for a bank or agency to describe a real migration. Would you trust an AI to rewrite the code that moves your money?