OpenAI has announced that its upcoming GPT-6 models deliver significantly better performance with legacy programming languages like COBOL, Fortran, and Visual Basic. Although these older languages have much less training data available compared to modern languages like Python, they still support vital global infrastructure, with IBM estimating that COBOL powers over 40 percent of online banking systems and the vast majority of ATM transactions.
To measure this progress, OpenAI created two specific evaluations. The MultiLangBench evaluation requires models to build a SQLite-compatible database engine from scratch across 20 languages. Using maximum reasoning effort, the GPT-6 Astra model achieves an average test pass rate of 87.4 percent, a major increase from the 38 percent scored by GPT-5.6 Terra. Meanwhile, the GPT-6 Luna model scores 40.3 percent while operating at a much lower cost.
On a separate COBOL migration benchmark designed to translate legacy COBOL code into modern languages like C++, Java, Kotlin, Python, or Rust, the GPT-6.1 Sol model reaches 86.6 percent accuracy at low reasoning effort, compared to just 5.4 percent for its predecessor. When reasoning effort is maximized, accuracy on this migration benchmark increases further to 97.9 percent.
This development matters because enhancing AI capabilities for legacy programming languages can drastically improve the maintenance and modernization of critical financial and administrative infrastructure worldwide.