OpenAI says its latest models have improved at handling legacy programming languages with limited training data, including COBOL, Fortran, and Visual Basic. These languages underpin critical systems in banking, government, and other industries, but have historically been difficult for AI models to work with because modern training data is scarce.
The company’s GPT-6 Astra model achieved an 87.4% pass rate on a benchmark test. The benchmark covered 20 languages, and OpenAI said the result was a substantial improvement over previous models.
The models are designed to help organizations maintain, modernize, and migrate essential legacy applications—a process that can be difficult and costly. They can better understand older code and translate it into newer languages while preserving its original business logic.