OpenAI says its latest AI models have improved at working with programming languages that have less code available for training, including COBOL, Fortran and Visual Basic. The company describes these as “low-resource” languages, though the source says they remain part of critical infrastructure worldwide.
OpenAI assessed the models with two evaluations: one asks them to build a database engine, while the other measures how well they translate COBOL into other languages. The results suggest that performance can differ substantially depending on the model and the amount of reasoning effort used.
Why older programming languages matter
Languages such as COBOL and Fortran have less available training code than Python, according to OpenAI. That scarcity can make them harder for AI models to handle reliably. The source gives no comparable detail about the amount of training code available for Visual Basic.
At the same time, the article points to the role of COBOL in financial infrastructure. IBM estimates that COBOL powers more than 40 percent of online banking systems and 95 percent of all ATM transactions. Those figures help explain why improving support for older code could matter beyond routine programming tasks.
For organizations that maintain software written in these languages, better model performance could be relevant to understanding or transforming existing code. The benchmark results, however, measure specific tasks; they do not establish how well a model would perform across every real-world system.
One test measures database building across languages
OpenAI built MultiLangBench to test whether models can create a SQLite-compatible database engine from scratch. The evaluation covers 20 languages and reports an average test pass rate.
With maximum reasoning effort, GPT-6 Astra achieved an average pass rate of 87.4 percent. The source compares that result with about 38 percent for GPT-5.6 Terra. GPT-6 Luna scored 40.3 percent, a smaller improvement over GPT-5.6 Terra’s 37.9 percent.
The comparison shows that results are not uniform across the latest models. GPT-6 Luna’s score was only slightly higher than GPT-5.6 Terra’s, while GPT-6 Astra’s was considerably higher under the stated maximum reasoning setting. The source also says GPT-6 Luna achieved its result at a much lower cost, without giving a cost figure.
COBOL translation shows a large reasoning effect
The second evaluation focuses specifically on COBOL migration. It asks models to translate COBOL into C++, Java, Kotlin, Python or Rust, and measures accuracy.
GPT-6.1 Sol reached 86.6 percent accuracy at low reasoning effort, compared with 5.4 percent for GPT-5.6 Sol. When reasoning effort was increased, GPT-6.1 Sol’s accuracy rose to 97.9 percent.
That gap makes reasoning effort a central part of the reported result. The low-effort and higher-effort scores describe different operating settings, so the headline accuracy depends on how the model is asked to work. The source does not provide further information about the benchmark’s test cases or how accuracy was calculated.
What the evaluations do and do not show
Together, the two evaluations offer evidence of progress on selected programming tasks involving multiple languages and COBOL translation. They also show why model name alone may not tell the whole story: the reported outcome changes with the task, the model and, in the migration test, the reasoning setting.
For teams considering AI assistance with legacy code, these figures are a benchmark signal rather than a guarantee of successful migration or dependable changes in production. The article reports test pass rates and translation accuracy, but it does not describe deployment results or the additional review a real software project might require.