After Anthropic disrupted IBM, OpenAI also posted a near-perfect score in COBOL migration.

Beating AI Express: In February this year, after Anthropic demonstrated Claude Code’s capability to modernize legacy COBOL systems, IBM’s stock price plummeted 13.2% in a single day. Eight months later, OpenAI released its latest results for the GPT-6 series on legacy programming languages, with a maximum success rate of 97.9% in COBOL code migration tests. COBOL, created in the 1950s, remains widely used in banking transactions, ATMs, and insurance systems. Many programs have operated for decades, their original developers long departed, and some business rules exist only within the code itself. Rewriting these into modern languages requires not just making the code run, but also ensuring fully consistent calculation results. For instance, COBOL’s fixed-point arithmetic discards excess decimal places unless explicitly ordered to round; a new program using rounding instead could lead to financial account errors. OpenAI specifically tested AI’s ability to convert COBOL programs to languages including C++, Java, Python, and Rust, requiring byte-for-byte identical output between the old and new programs. GPT-6.1 Sol achieved an 86.6% success rate at low inference settings, compared to just 5.4% for the previous GPT-5.6 Sol, with an estimated API cost reduction of 39%. Success rates climbed further to 97.9% when increasing inference settings. OpenAI also open-sourced MultiLangBench, which tasked AI with building SQLite-compatible database engines from scratch in 20 different programming languages. GPT-6 Astra posted an average test pass rate of 87.4%, versus 26.8% for GPT-5.6 Sol. Between the two models, their scores on COBOL were 72.1% and 10.9% respectively – COBOL ranks as the most challenging of the 20 tested languages.