OpenAI looks to help banks modernise legacy software their operations still depend on

OpenAI's artificial intelligence (AI) models have improved at working with legacy programming languages such as COBOL, Fortran and Visual Basic, helping companies maintain and modernise software that powers banking systems and other critical infrastructure.

The company said in an October 19 blog post that these languages have less code and documentation available for AI models to learn from than widely used languages such as Python and JavaScript.

“Maintaining and modernising this software can be difficult: documentation may be incomplete, original developers may have left the company, and important business rules may only exist in code,” OpenAI said.


According to figures cited by OpenAI, COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions, while Fortran is used in weather forecasting and nuclear reactor safety analysis.

The company said its latest models had improved both at building software in these languages and at moving existing programs to newer languages while preserving their behaviour. Its broader aim is to help models understand unfamiliar software projects, plan changes, use development tools, run tests and fix errors.

To test how well its models can write software in different languages, OpenAI asked them to build a database system from scratch. In this test, called MultiLangBench, GPT-6 Astra achieved an average pass rate of 87.4% across 20 programming languages, compared with 26.8% for GPT-5.6 Sol. The company said its models improved across all the languages tested, including COBOL, Fortran and Visual Basic.

OpenAI also tested how well its models could move existing COBOL programs to newer languages such as C++, Java, Kotlin, Python and Rust without changing how the software works.

OpenAI also tested how well its models could move existing COBOL programs to newer languages such as C++, Java, Kotlin, Python and Rust without changing how the software works.

GPT-6.1 Sol achieved 86.6% accuracy at a low reasoning setting, compared with 5.4% for GPT-5.6 Sol at the same setting. At a higher setting, GPT-6.1 Sol’s accuracy reached 97.9%, OpenAI said.

The company said this task went beyond translating code because the new version had to preserve the original program’s calculations and business rules. For example, COBOL can handle decimal calculations differently from other languages, and a small difference could affect financial results even if the translated program runs without errors.

“To test whether these behaviors are preserved, we evaluate our models on hidden tests that require byte-for-byte parity on observable outputs,” OpenAI said.

The company added that a migration is considered successful only if all tests pass. However, it noted that applying these capabilities to production software still requires expert review and testing against the system's business rules and dependencies.