In OpenAI's own tests, AI migrates COBOL code with 97.9% accuracy

IBM estimates that COBOL supports 95 percent of ATM transactions and 80 percent of in-person credit card transactions. For banks, this old code is not only a technical burden: it is often where the business rules live, and sometimes nowhere else. OpenAI now says its newest models can rewrite such programs in modern languages without changing how they behave, and reports 97.9 percent accuracy on its own migration test.

The difficulty is not only the age of the code. COBOL and Fortran, which is used in weather forecasting and nuclear reactor safety analysis, have far less training material available than Python or JavaScript, and their documentation is often incomplete. A rewritten program can run without a single error and still produce different numbers. COBOL, for example, discards extra decimal digits unless a program explicitly requests rounding, and in finance that detail is the whole calculation.

To score a migration, OpenAI's test translates existing COBOL programs into C++, Java, Kotlin, Python or Rust, and requires the new version to pass every hidden test, including checks that its observable outputs match the original. GPT-6.1 Sol reached 86.6 percent at low reasoning effort, against 5.4 percent for GPT-5.6 Sol, and 97.9 percent at higher effort. OpenAI also reported a 39 percent reduction in API costs at the lower setting.

The gains go beyond COBOL. On MultiLangBench, which asks models to build a database engine compatible with SQLite in 20 programming languages, GPT-6 Astra averaged 87.4 percent, up from 26.8 percent for GPT-5.6 Sol, and improved in every language. Its Fortran score rose from 21.4 to 83.5 percent, and COBOL, still the hardest language, climbed from 10.9 to 72.1 percent. The cheaper GPT-6 Luna averaged 40.3 percent, at approximately 78 percent less per sample: OpenAI estimates average API costs of $0.77 against $3.48.

What changes first is the cost of experimenting. For institutions that have long avoided touching core systems because replacing them is risky, a cheaper way to test rewrites is the practical point, and OpenAI argues that Luna's lower costs could let engineering teams run more trials without substantially increasing budgets. The limits are as clear as the gains: 97.9 percent is a score on OpenAI's own suite, not a guarantee, and the company itself points to COBOL as the weakest language in its results. Teams will still have to check migrated software against real workloads and undocumented dependencies before trusting it with critical operations.