OpenAI says its latest models can translate COBOL and other older languages into modern code with much higher benchmark accuracy, but experts warn that rewriting source files is only one part of replacing a critical legacy system.
On OpenAI’s evaluation, GPT-6.1 Sol preserved 87 percent of tested behavior at low reasoning effort and 98 percent at a higher setting while translating into languages including Java, Python, Rust and C++. The company counts a migration as successful only when all supplied tests pass. Anthropic is separately promoting Claude Code for legacy modernization, though it has not published comparable migration benchmarks.
Real systems are harder to define than benchmark tasks. Banks, tax agencies and other organizations may run code written in several languages, connect it to old data formats and depend on behavior that is poorly documented. Researchers interviewed by The Verge said OpenAI’s evaluation does not test the central work of dividing a large system into replaceable units or discovering implicit operational requirements.
Testing is therefore the bottleneck. Organizations can run old and new implementations in parallel, document translations and use formal verification for particularly important properties, but human review remains essential. Even a tool that correctly handles 99 percent of a vast codebase can leave a million scattered lines capable of subtle failures. The current progress may help maintenance and bounded migrations; it does not make wholesale COBOL replacement automatic.