OpenAI has released new results showing its latest models are much better at writing and migrating software in programming languages that have limited training data, such as COBOL, Fortran, and Visual Basic. These languages underpin critical systems: COBOL alone handles a large share of banking and ATM transactions, while Fortran is used in weather forecasting and nuclear safety.
To measure progress, OpenAI built two benchmarks. MultiLangBench asks models to build a SQLite-compatible database engine from scratch in 20 languages. The COBOL migration evaluation tests whether models can rewrite existing COBOL programs into other languages while preserving exact behavior, including subtle details like rounding rules. On MultiLangBench, GPT-6 Astra achieved an average pass rate of 87.4%, up from 26.8% for GPT-5.6 Sol, with gains in every language. COBOL remained the hardest, at 72.1%.
The cheaper GPT-6 Luna also improved over the previous model, reaching 40.3% at an estimated 78% lower API cost. On the COBOL migration test, GPT-6.1 Sol scored 86.6% at low reasoning effort, compared with 5.4% for GPT-5.6 Sol, at 39% lower cost; at higher effort it reached 97.9%. OpenAI notes that production use still requires expert review and testing against business rules.