OpenAI says new models sharply improve at COBOL coding

OpenAI said on Monday, October 19, 2026, that its newest models have improved sharply at writing and migrating code in "low-resource" programming languages such as COBOL, Fortran and Visual Basic. In OpenAI's own COBOL migration test, GPT-6.1 Sol scored 86.6% accuracy at low reasoning effort, compared with 5.4% for GPT-5.6 Sol.openai.com

OpenAI calls these languages low-resource because models have much less code and documentation to learn from than they do for Python or JavaScript. They still run core infrastructure, though. The company cited figures showing COBOL supports more than 40% of online banking systems and 95% of ATM transactions, and that Fortran is used in weather forecasting and nuclear reactor safety analysis. OpenAI said developers and enterprise customers had told it that coding models often struggle with these languages.openai.com

Two new evaluations

MultiLangBench, the first test, asks a model to build a SQLite-compatible database engine from scratch in each of 20 languages. The results are checked against the reference SQLite implementation using tests adapted from SQLite's own test suite. At maximum reasoning effort, GPT-6 Astra averaged an 87.4% test pass rate, up from 26.8% for GPT-5.6 Sol. Astra scored above 83% in 19 of the 20 languages. COBOL was the hardest, where Astra's score rose from 10.9% to 72.1%.openai.com

GPT-6 Luna, a cheaper model, averaged 40.3%. Its estimated API cost was $0.77 per sample, compared with $3.48 for GPT-5.6 Sol, or about 78% less. OpenAI is publishing the MultiLangBench setup, test suite and results for each language on GitHub.openai.com

Preserving old business logic

The second test asks models to convert COBOL programs into C++, Java, Kotlin, Python or Rust without changing what the programs do. Hidden tests require the outputs to match the original byte for byte, and a migration only counts if every test passes. OpenAI gave an example of why this is hard: COBOL's fixed-point arithmetic drops extra decimal digits unless rounding is requested. A rewrite that rounds instead can run without errors and still produce different financial results.openai.com

GPT-6.1 Sol's 86.6% score came at 61% of GPT-5.6 Sol's API cost. At higher reasoning effort, it reached 97.9%, which OpenAI said brings the test close to saturation.openai.com

Caveats and context

OpenAI designed and ran both tests itself, and no outside group has verified the results yet. The company also said that using these skills on production software "still requires expert review and testing against the system's business rules and dependencies".openai.com

Interest in COBOL has been growing. The language, created in 1959, climbed from 23rd place in October 2025 to 16th in the TIOBE popularity index, according to ITHardware. Companies still depend on these systems, and the pool of engineers who know how to maintain them is shrinking.ithardware.pl