OpenAI says its GPT-6 models now handle COBOL and Fortran far better

OpenAI published two new coding evaluations on October 19, 2026 aimed at older, less common programming languages such as COBOL, Fortran, OCaml and Visual Basic. There is far less public code in these languages for models to learn from, yet they still run a large share of banking, card payment and scientific software.

The first test, MultiLangBench, asks a model to build a SQLite-compatible database engine from scratch in one of 20 languages, then checks its output against the real SQLite. GPT-6 Astra averaged an 87.4% pass rate, up from 26.8% for the older GPT-5.6 Sol. COBOL rose from 10.9% to 72.1%, Fortran from 21.4% to 83.5% and Visual Basic from 27.5% to 89.4%. The cheaper GPT-6 Luna reached 40.3% at about 78% lower estimated cost than GPT-5.6 Sol.

The second test asks models to port COBOL programs to C++, Java, Kotlin, Python or Rust, and a port only counts if its outputs match the original byte for byte. GPT-6.1 Sol scored 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol, and 97.9% at higher effort. OpenAI is putting MultiLangBench and its results on GitHub, while warning that real migrations still need expert review.

Why it matters

Banks and governments sit on huge COBOL and Fortran codebases with few people left who know them, so models that can safely port that code could make long-delayed modernization projects realistic.

Source: OpenAI