The bet on ancient code
COBOL, a language born in 1959, remains embedded in much of the world's critical infrastructure, from tax systems to banks and airport services. According to data cited by IBM and reported by The Verge, it underpins more than 40 percent of online banking systems, 80 percent of card-present transactions and 95 percent of ATM operations. A problem no one dares tackle for fear of disrupting essential services.
Against this backdrop, OpenAI announced on October 19 improvements it described as substantial in its latest models' ability to migrate software written in obsolete languages to modern languages such as C++, Java, Kotlin, Python and Rust, without altering the program's original behavior. The company stresses that a migration is considered successful only if all tests are passed.
The numbers claimed by OpenAI
In internal tests, the GPT-6.1 Sol model achieved 87 percent accuracy in preserving behavior with low reasoning effort, rising to 98 percent by increasing the number of analysis steps. In the same scenario, GPT-6 Astra stopped at 86 percent, GPT-6 Sol at 21 percent and GPT-5.6 Sol at 5 percent. OpenAI also reported improvements in the models' ability to build database software in various languages, including COBOL.
Anthropic also published an operating manual for code modernization in February, proposing its Claude Code agent as a tool to migrate COBOL applications to Java or Python. The company did not provide performance benchmarks, but argued that the economics of modernization have changed. The Verge reports that shares in IBM, which derives part of its revenue from maintaining legacy systems, fell sharply after the announcement.
The researchers' reservations
The experts interviewed by The Verge acknowledge progress but warn about the gap between laboratory tests and the safe replacement of critical systems. Vadim Zaytsev of the University of Twente notes that rewriting code is not the same as replacing a system, because multiple languages, data formats, technical integrations and organizational expectations come into play. Arie van Deursen of Delft University calls OpenAI's post superficial, noting that the tests involve simple tasks close to those seen in training and that maintaining a system is much easier than modernizing it.
Alastair Donaldson of Imperial College London adds that the main difficulty is not rewriting the code but understanding how the system works, a gap common in legacy software and worsened by the shortage of COBOL programmers. Jacopo Mauro of the University of Southern Denmark points to formal verification as a possible automated tool, but warns that applying it to legacy code would require enormous work and that, until then, human verification remains essential. An OpenAI spokesperson admitted that the problem is far from solved, while Anthropic declined to comment. It therefore remains to be seen whether the claimed results can be replicated on real, complex systems.
