OpenAI GPT-6.1 Sol achieves 97.9% success rate in COBOL migration

OpenAI has released test results showing its mid-range model, GPT-6.1 Sol, can migrate COBOL programs to modern languages such as Java, Rust, or Python with 97.9% accuracy without altering behavior. This represents a significant leap from the previous generation, GPT-5.6 Sol, which achieved only 5.4% accuracy under similar conditions when using reduced reasoning time.

The model's performance is particularly notable for "low-resource" languages like COBOL, Fortran, and Visual Basic, which have less available documentation and code for training compared to Python or JavaScript. These legacy languages remain critical to global infrastructure; COBOL supports over 40% of online banking systems, 80% of in-store credit card payments, and 95% of ATM transactions. Fortran continues to be used for weather forecasting and nuclear reactor safety analyses.

To evaluate these capabilities, OpenAI utilized its MultiLangBench benchmark, which requires models to build SQLite-compatible database engines across 20 programming languages. In these tests, GPT-6.1 Sol showed substantial improvements across several languages, including Fortran (83.5%) and Visual Basic (89.4%).