The Quiet Revolution of OpenAI: New Models Breaking the COBOL and Legacy Code Barrier
While most discourse around artificial intelligence focuses on modern languages like Python or JavaScript, a technological heart of the past still beats beneath the surface of the world's largest financial and governmental systems. Languages like COBOL, Fortran, and Visual Basic are the backbone of over 40% of online banking systems and 95% of ATM transactions. The problem? These are "low-resource" languages—with little documentation, few active programmers, and immense difficulty in maintenance. OpenAI recently announced a significant leap in its models' ability to handle this code, which may change the way organizations modernize their core systems [1, 4, 7].
Why is this critical?
The great challenge in COBOL is not just writing new code, but migrating existing systems to modern languages (like Java, Python, or Rust) without breaking business logic. Imagine a banking system that calculates interest or rounds numbers in a specific way—any slight deviation in the result, even at the last digit, could lead to operational collapse. Until today, language models struggled to maintain "byte-for-byte parity" in such migration processes [3, 7, 8].
The numbers behind the technological breakthrough
In a series of tests conducted by OpenAI, the new models demonstrated impressive performance that leaves previous generations behind:
- •The MultiLangBench Test: The GPT-6 Astra model reached an average success score of 87.4% in building software in 20 different languages, compared to only 26.8% in the GPT-5.6 Sol model [4, 7].
- •COBOL Migration: Here, the most impressive success was recorded. The GPT-6.1 Sol model reached 86.6% accuracy with low computational effort, and jumped to 97.9% accuracy with high computational effort. This is a dramatic improvement over past results, bringing these systems to a level of maturity where one can start considering practical use [2, 3, 8].
- •Economic Efficiency: Beyond accuracy, OpenAI notes that the new models significantly reduce API costs, making large-scale code migration projects much more economical than in the past [2, 7, 8].
What does this mean for us?
It is important to keep our feet on the ground. OpenAI emphasizes that despite the excellent results, migrating critical systems (Production) still requires careful human supervision. Artificial intelligence can perform the "grunt work" of translating logic and code, but the responsibility for verifying business rules and complex system dependencies remains with experts [3, 7, 8].
Summary and conclusions
The ability of models like GPT-6.1 Sol to handle "forgotten" languages is real news for organizations that feel trapped by technical debt. Instead of replacing entire high-risk systems, we are approaching an era where AI enables "live translation" of core systems, while maintaining their exact behavior. For developers and organizations, it is time to start evaluating these tools as part of the organization's modernization strategy.
Sources and further information: