OpenAI says GPT-6 is much better with legacy programming languages such as COBOL and Fortran

OpenAI has made its latest AI models much better at working with older programming languages such as COBOL, Fortran, and Visual Basic. There is much less code available to train AI on these languages than there is for languages like Python. But according to OpenAI, these older languages still run important computer systems around the world. IBM estimates that COBOL runs more than 40 percent of online banking systems and handles 95 percent of all ATM transactions.

OpenAI created two tests to measure how well its models have improved. MultiLangBench (GitHub) tests whether a model can build a database engine from scratch that works with SQLite. When using its highest reasoning setting, GPT-6 Astra passes an average of 87.4 percent of the tests across 20 programming languages. GPT-5.6 Terra passes about 38 percent. GPT-6 Luna scores 40.3 percent, just above GPT-5.6 Terra's 37.9 percent. However, GPT-6 Luna costs much less to use.

Another test measures how accurately AI can convert COBOL code into C++, Java, Kotlin, Python, or Rust. On this test, GPT-6.1 Sol reaches 86.6 percent accuracy when using a low reasoning setting. GPT-5.6 Sol reaches only 5.4 percent. When GPT-6.1 Sol uses a higher reasoning setting, its accuracy rises to 97.9 percent.