OpenAI GPT-6 Astra Reaches 72% Accuracy on COBOL

OpenAI has published evaluation results for its latest artificial intelligence models on legacy and low-resource programming languages, alongside the release of MultiLangBench on GitHub. The new GPT-6 Astra model scored 87.4% overall across MultiLangBench's 20 languages—a steep increase from the 26.8% achieved by GPT-5.6 Sol. On the benchmark's hardest task, building a SQLite-compatible database engine in COBOL, GPT-6 Astra reached 72.1% accuracy, exceeding an 83% pass rate in the other 19 languages, which include Fortran, OCaml, and Zig.

For code conversion tasks, GPT-6.1 Sol achieved a 97.9% accuracy rate on COBOL-to-modern-language migration when configured for high reasoning effort. Even under low reasoning effort, automated migration accuracy jumped from 5.4% to 86.6% while dropping API costs by 39%. OpenAI's strict evaluation protocol demands byte-for-byte output parity against reference database files. This precise standard catches subtle issues like fixed-point arithmetic errors, which can otherwise silently break interest or settlement calculations in ported banking applications.

These advances target mission-critical systems where original maintainers are long gone. COBOL underpins over 40% of online banking systems, 80% of in-person credit card swipes, and 95% of ATM transactions, while Fortran handles weather forecasting and nuclear safety analysis. While estimated API usage costs are down, practitioners must still invest in test creation, dependency mapping, and thorough code reviews. Because edge-case errors can still slip through to reconciliation procedures, developers are advised to use these models primarily for automated code extraction, test drafting, and assisted migration paired with rigorous differential testing.