OpenAI advances coding in COBOL, Fortran and more

What low-resource means here

OpenAI means languages with relatively little code and documentation for models to learn from, such as COBOL, Fortran and Visual Basic. Those languages still run critical systems: COBOL supports more than 40% of online banking, 80% of in-person card transactions and 95% of ATM transactions, while Fortran shows up in weather forecasting and nuclear reactor safety analysis (OpenAI).

MultiLangBench: SQLite across 20 languages

MultiLangBench asks a model to build a SQLite-compatible database engine in a chosen language. Database files are the shared interface: the candidate gets an initial database and a SQL query, then returns a modified database. The reference SQLite implementation checks the result. Tests come from SQLite's own suite.

At maximum reasoning effort, GPT-6 Astra reaches an average 87.4% test pass rate across the 20 languages, up from 26.8% for GPT-5.6 Sol. Gains include COBOL (10.9% to 72.1%), OCaml (25.6% to 85.8%), Fortran (21.4% to 83.5%) and Visual Basic (27.5% to 89.4%). Astra scores above 83% in 19 languages; COBOL remains hardest at 72%.

GPT-6 Luna is the lower-cost option: 40.3% on average, better than GPT-5.6 Sol in every language, at about 78% lower estimated API cost ($0.77 per sample versus $3.48). OpenAI is releasing the evaluation setup and test suite on GitHub (MultiLangBench).

COBOL migration with byte-level parity

The second eval migrates COBOL programs to C++, Java, Kotlin, Python or Rust. Success requires byte-for-byte parity on hidden tests. COBOL's fixed-point arithmetic, for example, discards excess decimals unless rounding is requested. A rewrite that rounds instead can run without errors while changing financial results.

At low reasoning effort GPT-6.1 Sol reaches 86.6% accuracy, versus 5.4% for GPT-5.6 Sol, at about 39% lower cost. At higher reasoning effort accuracy reaches 97.9%. OpenAI still expects expert review and tests against business rules before production migrations.

Why it matters

Teams modernizing sparse-language codebases often face incomplete docs and business rules that live only in code. The new numbers show large accuracy gains at lower cost. Benchmark on your own COBOL or Fortran samples before putting a model in a migration pipeline.