OpenAI targets COBOL migration projects
In brief
- OpenAI claims its latest models migrate COBOL projects with 86.6 % accuracy, up from 5.4 % in previous versions.
- Internal tests like MultiLang evaluate software creation and migration across 20 languages, including COBOL and Fortran.
- The company stresses the need for human verification despite progress, due to opaque evaluations and production risks.
OpenAI claims significant progress in migrating software written in COBOL and other legacy languages, dubbed « low-resource languages ». The company states that its latest models, such as Astra and GPT-6, can now generate or migrate programs in these languages with higher accuracy. Internal tests, including MultiLang, assess the ability to create database engines in 20 languages or migrate COBOL projects to C++, Java, Kotlin, Python, or Rust.
OpenAI highlights specific challenges with these languages: incomplete documentation, lack of trained developers, or business rules embedded in the code. Evaluations show notable improvements, with GPT-6.1 Sol achieving 86.6 % success in COBOL migration, compared to 5.4 % for GPT-5.6 Sol. Scores improve further with increased reasoning effort, but OpenAI emphasizes the need for human verification before production use.
The announcement lacks transparency: details of the evaluations, such as the number of tested programs or their origin, are not public. OpenAI acknowledges that resources for these languages are scarce compared to projects like SQLite. The community cannot reproduce the tests, limiting the credibility of the results.
What it changes for you
- Finance & BankingBanks could accelerate the migration of their COBOL systems to modern languages, reducing maintenance costs and risks linked to the lack of trained developers.
- Public SectorGovernments using COBOL for critical transactions could consider partial migrations, subject to thorough human verification.
This summary gives you the essentials. The full article, its figures and quotes are on Next.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error