After Anthropic shook IBM, OpenAI also delivered a near-perfect COBOL migration score.

Beating AI News Flash: In February of this year, after Anthropic promoted Claude Code's ability to transform legacy COBOL systems, IBM's stock price plummeted 13.2% in a single day.

Eight months later, OpenAI also announced the latest results of the GPT-6 series on legacy programming languages, with the highest success rate in COBOL code migration tests reaching 97.9%.

COBOL was born in the 1950s and is still widely used in bank transactions, ATMs, and insurance systems. Many programs have been running for decades, their original developers long gone, and some business rules exist only in the code. Rewriting them into modern languages requires not only making the code run but also ensuring the computation results are exactly identical. For example, in fixed-point arithmetic, COBOL directly truncates extra decimal places if rounding is not explicitly required. If a new program switches to rounding, it could cause errors in financial accounts.

OpenAI specifically tested AI's ability to rewrite COBOL programs into C++, Java, Python, Rust, and other languages, requiring the old and new programs to produce byte-for-byte identical outputs in tests. GPT-6.1 Sol achieved an 86.6% success rate at the low reasoning tier, compared to only 5.4% for the previous generation GPT-5.6 Sol, with estimated API costs also reduced by 39%. After raising the reasoning tier, the success rate further reached 97.9%.

OpenAI also open-sourced MultiLangBench, requiring AI to write a SQLite-compatible database engine from scratch in 20 different programming languages. GPT-6 Astra achieved an average test pass rate of 87.4%, while GPT-5.6 Sol reached 26.8%. Among these, the two models scored 72.1% and 10.9% respectively on COBOL, which was the hardest of the 20 languages.