AI improves COBOL and Fortran programming

OpenAI says its latest models are significantly better at programming and migrating software written in languages with limited available data, such as COBOL, Fortran, OCaml and Visual Basic. These technologies are old, but they still support banks, ATMs, weather forecasting systems and other essential services.

The problem is that these languages have less publicly available code and documentation than Python or JavaScript. Many companies also depend on programs created decades ago, with incomplete documentation and business rules explained only inside the code itself.

That makes modernizing them risky. A migration can appear correct and still change an important financial rule, such as how COBOL discards decimals when explicit rounding is not requested.

Better results when building software from scratch

OpenAI measured progress with MultiLangBench, an evaluation in which models must build a SQLite-compatible engine from scratch in 20 languages. The system has to organize the project, manage data, handle errors and add features without breaking what already works.

In this test, GPT-6 Astra achieved an average pass rate of 87.4%, compared with 26.8% for GPT-5.6 Sol. The model improved in all 20 languages analyzed.

The results were especially notable in several less common languages:

• COBOL: from 10.9% to 72.1%
• OCaml: from 25.6% to 85.8%
• Fortran: from 21.4% to 83.5%
• Visual Basic: from 27.5% to 89.4%

Astra exceeded 83% in 19 of the 20 languages. COBOL remained the most difficult, at 72%.

There was also a more affordable option. GPT-6 Luna reached an average of 40.3% and improved on its predecessor in every language tested, at an estimated cost of $0.77 per sample, compared with $3.48. That represents an approximate reduction of 78%.

Migrating COBOL without changing its behavior

The second evaluation examined whether a model could convert COBOL programs to C++, Java, Kotlin, Python or Rust while preserving exactly what they originally did.

To be considered correct, the migration had to pass hidden tests and produce identical results, byte for byte. It was not enough for the new program to run: it had to preserve rules such as rounding, the handling of missing records and transaction management.

In this test, GPT-6.1 Sol achieved 86.6% accuracy with a low reasoning level, compared with 5.4% for GPT-5.6 Sol. It also cost 39% less. With a high reasoning level, it reached 97.9%.

What changes for companies

For a team maintaining legacy systems, these advances could reduce the work involved in finding bugs, understanding unfamiliar repositories or preparing a migration. For example, a company could ask for help moving a COBOL program to Java without losing the logic used to calculate payments or process banking records.

But the results come from controlled evaluations. Before using a migration in production, experts still need to review the code, external dependencies and business-specific rules. A small error can alter a balance, a payroll calculation or a regulatory report.

OpenAI has also published the MultiLangBench environment and tests so developers can review how performance is measured and identify the most important gaps for their own projects.

The important signal is that AI is becoming more useful for software companies cannot replace overnight. The next point to watch is whether these results hold up in real systems, which are much larger and have less visible rules than those in an automated test.