COBOL still supports systems used in banking, government, and other critical services. New AI results suggest that models may be getting better at translating this old code into modern languages. That could help organizations maintain software that is difficult to change, but a successful translation alone does not make an entire system safe to replace.
Why organizations still depend on COBOL
COBOL, short for “common business-oriented language,” dates to 1959. Its age has not made it easy to retire: the language remains part of systems that organizations depend on, including online banking, credit card transactions, and ATM transactions.
IBM says COBOL is the foundation for more than 40 percent of online banking systems, supports 80 percent of in-person credit card transactions, and handles 95 percent of ATM transactions. The source also describes its use in government systems, including Social Security, the IRS, and services run by most states.
Replacing software in those settings carries a practical risk: the existing system may be old and hard to understand, but it is already doing work that people and organizations rely on. The article notes that many programmers familiar with legacy languages have retired or died. In other fields, languages such as Fortran also remain in use for work including scientific computing, weather forecasting, atmospheric and ocean studies, and codes critical to nuclear security.
What the latest AI results show
OpenAI reported that its latest models had improved at migrating older software into modern programming languages while preserving what the original program does. Its examples included moving COBOL code into C++, Java, Kotlin, Python, and Rust. The company says a migration counts as successful only if all tests pass.
In OpenAI’s own tests, GPT-6.1 Sol achieved 87 percent behavior preservation accuracy at low reasoning effort and 98 percent at higher reasoning. At low reasoning effort, GPT-6 Astra scored around 86 percent, GPT-6 Sol scored 21 percent, and GPT-5.6 Sol scored 5 percent. OpenAI also reported progress in its models’ ability to build database software in various languages, including COBOL.
Those figures offer a measure of how well models performed on the tasks OpenAI evaluated. They do not, by themselves, establish that a business or government can replace a working legacy system safely. OpenAI spokesperson Oscar Haines said the problem is “still far from a solved problem,” and that the company is exploring ways to improve tests and models’ understanding of programming languages.
Translating code is only one part of modernization
A legacy system is more than the code that can be translated. Vadim Zaytsev, an associate professor of software evolution at the University of Twente in the Netherlands, pointed to data stored in different formats, software written in multiple languages, connections to other systems, and the expectations of the people and organizations using them.
That is why a system-level migration involves decisions beyond whether the translated code behaves like a particular piece of original code. The work also requires figuring out how a large system can be divided into smaller parts that can be replaced individually. Zaytsev said AI is not currently playing a major role in replacing or migrating legacy COBOL systems.
Arie van Deursen, a software engineering professor at Delft University of Technology in the Netherlands, argued that OpenAI’s evaluations do not capture that complexity. He described the company’s blog as “superficial,” saying it gave few details about the migration task and relied on a relatively simple, well-defined benchmark. In his view, the test was far from the work of modernizing real systems.
The difference is consequential for organizations weighing AI tools. Maintaining an aging system and modernizing it are distinct tasks: a tool may help with a bounded coding problem without resolving the broader challenge of changing the system around it.
Progress does not yet mean safe replacement
Anthropic has also promoted AI for legacy code. In February, it published a “code modernization playbook” and presented Claude Code as a way to migrate COBOL into languages such as Java or Python. The article says Anthropic did not provide performance benchmarks for those migrations.
Experts who spoke to The Verge agreed that AI tools have made progress with legacy code, while cautioning that the results do not show that critical systems can now be safely replaced. The central test is not just whether a model can rewrite a portion of code; it is whether teams can understand, separate, and replace parts of a complicated system while accounting for its data, connections, and users.
For now, AI’s results may offer a reason to keep exploring code migration, but they leave the hardest system-level questions open. Organizations that depend on COBOL still need evidence that a proposed change preserves the behavior and relationships their systems rely on, not just a promising score on a limited evaluation.