A great deal of the world's critical software is written in languages that AI coding tools handle badly. COBOL, Fortran and Visual Basic have far less public code and documentation to learn from than Python or JavaScript, yet OpenAI notes that COBOL alone supports more than 40% of online banking systems. OpenAI now reports large gains on two tests it built to measure this, and the numbers matter to banks, insurers and governments sitting on decades-old code.
The first test, MultiLangBench, asks a model to build a SQLite-compatible database engine from scratch in one of 20 languages, checked against tests adapted from SQLite's own suite. At maximum reasoning effort, GPT-6 Astra passes 87.4% of tests on average, against 26.8% for GPT-5.6 Sol, and scores above 83% in 19 of the 20 languages. COBOL is still the hardest at 72.1%, up from 10.9%; Fortran rises from 21.4% to 83.5% and Visual Basic from 27.5% to 89.4%. The cheaper GPT-6 Luna reaches 40.3% at about $0.77 per sample, against $3.48 for GPT-5.6 Sol.
The second test is closer to real work: migrating existing COBOL programs to C++, Java, Kotlin, Python or Rust, where a migration only counts if hidden tests show byte-for-byte identical output. That standard matters because COBOL's fixed-point arithmetic drops extra decimal digits unless told to round, and a rewrite that rounds can run cleanly while producing different results. GPT-6.1 Sol reaches 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol at the same setting, and 97.9% at higher effort, close to the ceiling.
For readers weighing coding models, this is the strongest public claim yet that frontier models can help with legacy languages, and OpenAI has released the MultiLangBench setup and per-language results on GitHub, which lets others check it. That openness counts for more than the headline figures.
The caveats are significant. Both tests were designed by OpenAI and report only OpenAI models, so there is no comparison with Claude or Gemini. Building a database engine to a known specification is not the same as maintaining a 40-year-old banking system whose business rules may exist only in the code, written by developers who have long since left. Near-saturation on the migration test says more about the test than about production readiness, and OpenAI itself says real migrations still need expert review and testing against business rules. Careful pilots on a team's own code remain the only reliable measure.