OpenAI improves GPT-6 handling of legacy code like COBOL

OpenAI has optimized its latest AI models to handle under-resourced legacy programming languages such as COBOL, Fortran, and Visual Basic. These classic syntaxes suffer from far smaller pools of available training code compared to ubiquitous modern languages like Python, yet they continue to anchor vital global infrastructure. According to IBM estimates, COBOL alone currently powers more than 40 percent of online banking systems and 95 percent of all ATM transactions globally.

To track progress in these specialized domains, OpenAI designed MultiLangBench, an evaluation that challenges models to build a complete SQLite-compatible database engine from scratch across 20 distinct programming languages. When configured with maximum reasoning effort, GPT-6 Astra achieved an average test pass rate of 87.4 percent, climbing from approximately 38 percent recorded by GPT-5.6 Terra (which scored 37.9 percent). Additionally, GPT-6.1 Sol reached about 87 percent on the same suite, while the lower-cost GPT-6 Luna posted a 40.3 percent pass rate.

The improvements extend directly into code modernization via OpenAI's COBOL migration benchmark, which measures translation from COBOL into modern targets like C++, Java, Kotlin, Python, or Rust. At low reasoning effort, GPT-6.1 Sol reached 86.6 percent translation accuracy, standing in stark contrast to the 5.4 percent accuracy achieved by GPT-5.6 Sol. Increasing the reasoning effort pushed GPT-6.1 Sol's translation accuracy to 97.9 percent.

For software practitioners, system architects, and enterprise IT teams, these metrics mark a critical shift in automated refactoring. Translating legacy mainframes historically carried high failure rates and required specialized manual labor due to scarce expertise. With translation accuracy reaching nearly 98 percent, developers can automate bulk migrations into modern languages with dramatically reduced debugging overhead and lower overall execution costs.