Thông tin chính
On October 19, OpenAI announced that its GPT-6 series models have achieved dramatic improvements in generating and translating code for legacy languages like COBOL, Fortran, and Visual Basic. These languages are often categorized as low-resource, as they lack the extensive training data available for modern languages like Python.
In migration tests where COBOL was translated into languages such as C++, Java, or Rust, the GPT-6.1 Sol model reached an 86.6% success rate, increasing to 97.9% with higher reasoning settings. This marks a substantial improvement over the 5.4% success rate recorded by the previous GPT-5.6 Sol model. OpenAI noted that these tests required byte-level output accuracy to pass.
Despite these advancements, OpenAI emphasizes that human oversight remains essential for production systems. While the model can handle complex logic, the company advises that expert reviews and rigorous testing against business rules are still necessary before deployment. COBOL remains the most challenging language, as it is the only one that did not exceed an 83% pass rate in the MultiLangBench evaluations.
This development offers a potential solution for maintaining critical infrastructure, such as banking systems, where legacy code persists despite a shrinking pool of skilled engineers. OpenAI developed these capabilities in response to feedback from developers and corporate clients struggling with the limitations of existing AI tools in legacy environments.
