OpenAI published on October 19, 2026 evaluation results showing significant advances by its latest models in low-resource programming languages such as COBOL, Fortran, OCaml and Visual Basic. These languages have less code and documentation available for training than Python or JavaScript, yet they support critical infrastructure worldwide.
According to the company’s official post, COBOL still supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM operations. Fortran continues to be used in weather forecasting and nuclear reactor safety analysis. Improving AI’s ability to work in these languages can help teams maintain and modernize essential systems.
The tests
OpenAI developed two complementary evaluations:
- MultiLangBench — asks the model to build a SQLite-compatible database engine from scratch in 20 different languages. Measures test pass rate and API cost.
- COBOL migration evaluation — tests whether the model can migrate COBOL programs to C++, Java, Kotlin, Python or Rust while preserving original behavior, including rounding rules and record handling.
On MultiLangBench, GPT-6 Astra achieved an average 87.4% test pass rate across 20 languages, compared with 26.8% for GPT-5.6 Sol. Improvements occurred in every language, including COBOL (from 10.9% to 72.1%), Fortran (21.4% to 83.5%) and Visual Basic (27.5% to 89.4%). The cheaper GPT-6 Luna reached 40.3% with roughly 78% lower API cost.
In COBOL migration, GPT-6.1 Sol reached 86.6% accuracy at low reasoning effort (versus 5.4% for the previous model) and 97.9% at higher effort. OpenAI released the MultiLangBench setup on GitHub so developers can inspect the tests.
Why it matters
Maintaining and modernizing legacy software is a real problem for banks, governments and companies. Incomplete documentation, departed original developers and business rules that exist only in code make the work difficult. Advances in behavior-preserving migration can reduce risk and cost, although the company makes clear that production use still requires expert review and testing against the system’s business rules.
The results come from OpenAI’s internal evaluations. They indicate real progress on a problem the industry has faced for decades, but they do not eliminate the need for human validation in critical environments.
Sources
Transparency: This content was created, edited, or reviewed with the aid of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
