On October 20, OpenAI released a research report stating that it has significantly enhanced the build and migration capabilities of models such as GPT-6 Astra on "low-resource programming languages" like COBOL, Fortran, and Visual Basic. These languages have scarce public code and documentation, yet they form the backbone of critical systems such as banking, ATMs, weather forecasting, and even nuclear safety analysis—rarely touched in normal times, but with the power to shake everything when things go wrong.
To quantify the capability, OpenAI set two suites of exams: MultiLangBench requires the model to build from scratch a SQLite-compatible database engine using 20 languages; the other requires migrating COBOL programs to C++, Java, Python, and Rust, while preserving the original business behavior intact. The report card is impressive: GPT-6 Astra achieved an average test pass rate of 87.4% across 20 languages, more than triple the previous generation GPT-5.6 Sol's 26.8%; looking only at the niche languages, COBOL leaped from 10.9% to 72.1%, Fortran rose from 21.4% to 83.5%, and Visual Basic jumped from 27.5% to 89.4%.
The budget tier can compete too, with migration approaching the evaluation ceiling
Even more pleasing is the cost side. The cheaper GPT-6 Luna has an average pass rate of 40.3%, but its API costs are about 78% lower than GPT-5.6 Sol's. In the COBOL-specific migration, GPT-6.1 Sol achieved an accuracy of 86.6% even at low reasoning effort, far outstripping GPT-5.6 Sol's 5.4%, while cutting costs by 39%; at higher reasoning effort, accuracy hit 97.9%, nearly touching the evaluation ceiling—meaning that when replacing decades-old bank core code with a modern language, machines can already deliver an almost usable answer.
OpenAI noted that these advances can help enterprises maintain legacy systems, fix defects, and advance software modernization. But the company also added a sobering footnote: moving AI migration results into production still requires expert review, combined with testing against specific business rules and dependencies. After all, the code underpinning ATMs and nuclear power plants tolerates no "close enough"—models can greatly shorten the distance from understanding to rewriting, but they are still far from the point of making the final call for humans.