OpenAI's GPT-6 series showed sharply improved coding performance over earlier models in programming languages with limited training data, including COBOL, Fortran and Visual Basic. It is expected to be used for maintaining and modernising old software used in banks and major industrial facilities.

On Oct. 19, OpenAI released results from an evaluation of GPT-6 models' ability to handle low-resource programming languages. OpenAI developed a benchmark called MultiLangBench to measure performance. The evaluation tests how well an AI model builds an SQLite-compatible database engine in 20 languages including C, Python and COBOL, and how many set functional tests it passes.

In the experiments, GPT-6 Astra posted an average test pass rate of 87.39 percent, far above 26.77 percent for the previous model, GPT-5.6 Sol. GPT-6.1 Sol also recorded 87.04 percent.

Improvements were also clear by language. GPT-6 Astra's COBOL test pass rate rose to 72.1 percent from 10.9 percent, while Fortran climbed to 83.5 percent from 21.4 percent. Visual Basic also increased to 89.4 percent from 27.5 percent.

The low-cost GPT-6 Luna model also exceeded GPT-5.6 Sol, with an average of 40.31 percent. The estimated API cost per test was $0.77, about 1,020 won, about 78 percent lower than the previous model.

Better performance was also confirmed in tests converting existing COBOL programs into modern languages such as C++, Java and Python. GPT-6.1 Sol recorded accuracy of 86.6 percent at a low reasoning level and 97.9 percent at a high reasoning level.

COBOL is still used in financial transactions and bank computer systems, and Fortran is also used in weather forecasting and in science and technology fields.

OpenAI explained that the results could support maintenance of old core software and system transitions. It said the figures were obtained from an evaluation it designed in-house. It added that expert review and further validation would be needed before applying the models to real financial and industrial systems.