OpenAI reports that GPT-6 Astra shows significant improvements on low-resource languages such as COBOL and Fortran, achieving an average pass rate of 87.4% on MultiLangBench—three times higher than its predecessor. Pass rates for COBOL, Fortran, and Visual Basic are 72.1%, 83.5%, and 89.4%, respectively. The low-cost GPT-6 Luna API reduces costs by 78%. While it can accelerate code migration, production deployment still requires expert review to ensure security.

Author and source: AIBase

On October 20, OpenAI released a research report stating that it has significantly improved the ability of models such as GPT-6 Astra to build and migrate code in "low-resource programming languages" like COBOL, Fortran, and Visual Basic. These languages have minimal publicly available code and documentation, yet they form the backbone of critical systems—including banking, ATMs, weather forecasting, and even nuclear safety analysis—systems that are rarely touched under normal circumstances but can have system-wide consequences when they fail.

To quantify its capabilities, OpenAI set two challenging benchmarks: MultiLangBench required the model to build a SQLite-compatible database engine from scratch in 20 languages, while the other task demanded migrating COBOL programs to C++, Java, Python, and Rust while preserving the original business behavior exactly. The results were impressive: GPT-6 Astra achieved an average pass rate of 87.4% across all 20 languages—more than triple GPT-5.6 Sol’s 26.8%. For less common languages, improvements were even more dramatic: COBOL rose from 10.9% to 72.1%, Fortran from 21.4% to 83.5%, and Visual Basic from 27.5% to 89.4%.

Even budget tiers can compete, as migration approaches the evaluation ceiling.

Even more appealing is the cost side. The cheaper GPT-6 Luna achieves an average pass rate of 40.3%, while reducing API costs by approximately 78% compared to GPT-5.6 Sol. In COBOL-specific migrations, GPT-6.1 Sol reaches an accuracy of 86.6% even under low inference intensity—far surpassing GPT-5.6 Sol’s 5.4%—while cutting costs by 39%. When switching to higher inference intensity, accuracy jumps to 97.9%, nearly hitting the evaluation ceiling—meaning machines can now deliver nearly production-ready results when modernizing decades-old banking core code.

OpenAI noted that this advancement can help businesses maintain legacy systems, fix bugs, and drive software modernization. However, the company also offered a sober caveat: migrating AI-generated changes into production environments still requires expert review and testing against specific business rules and dependencies. After all, the code supporting ATMs and nuclear power plants cannot tolerate even a single "close enough"—while models can significantly shorten the gap between understanding and rewriting, they are still far from replacing human decision-making.