OpenAI said on October 19 that its newest artificial intelligence models have made substantial progress converting software written in older programming languages into modern code, an advance the company says could help organizations maintain and update aging systems. COBOL, a language launched in 1959, is among those OpenAI highlighted. The company reported that its GPT-6.1 Sol model reached 87 percent behavior preservation accuracy at low reasoning effort and 98 percent at higher reasoning when migrating COBOL code to languages including C++, Java, Kotlin, Python and Rust without changing what the original software does. OpenAI said a migration counts as successful only if all tests pass.
The results matter because COBOL remains deeply embedded in US critical infrastructure. According to IBM, the language underlies more than 40 percent of online banking systems, supports 80 percent of in-person credit card transactions and handles 95 percent of ATM transactions, and it also runs through government systems from Social Security to the IRS to services operated by most states. One rough estimate puts the annual cost of COBOL's inefficiencies to US GDP at around $105 billion. The same difficulties apply to other legacy languages such as Fortran, still used in scientific computing, weather forecasting and codes critical to nuclear security.
In OpenAI's tests, GPT-6 Astra scored around 86 percent at low reasoning effort, GPT-6 Sol 21 percent and GPT-5.6 Sol 5 percent. OpenAI also reported gains in its models' ability to build database software in various languages, including COBOL. Anthropic has moved on the same territory: in February it published a step-by-step code modernization playbook and pitched its coding agent, Claude Code, as a way to move COBOL into languages like Java or Python, offering no performance benchmarks. Shares in IBM, which earns part of its revenue maintaining legacy systems, fell sharply after that announcement.
Experts told The Verge that AI tools have improved at handling legacy code but remain far from safely replacing critical systems. Vadim Zaytsev, an associate professor of software evolution at the University of Twente, said rewriting code is not the same as replacing a system, citing multiple languages, stored data, technical integrations and organizational expectations as complicating factors. He said AI is not playing a major role at the moment in replacing or migrating legacy COBOL systems.
Arie van Deursen, a software engineering professor at Delft University of Technology, called OpenAI's blog post superficial, saying its benchmark tested a relatively simple, well-defined task and that maintaining an aging system is far easier than modernizing one. Alastair Donaldson, a professor of programming languages at Imperial College London, said the main challenge is not understanding how a system works, a common problem with legacy code, and that nobody in their right mind would use COBOL for a new development project. OpenAI spokesperson Oscar Haines acknowledged to The Verge that legacy code is still far from a solved problem. Anthropic did not comment in time for publication.
More AI news from TechManNews.
