Can GPT-6 Finally Fix COBOL? Inside OpenAI's 97.9% Legacy Leap

According to reports from The Decoder, OpenAI's latest GPT-6 series brings massive improvements to low-resource programming languages like COBOL, Fortran, and Visual Basic. While mainstream models excelled at Python and JavaScript, they historically struggled with the decades-old infrastructure powering global banking systems.

What do the benchmarks reveal?

Data from OpenAI's new MultiLangBench shows GPT-6 Astra hitting an 87.4 percent test pass rate across 20 languages, a huge jump from the 38 percent achieved by GPT-5.6 Terra. More importantly, in dedicated COBOL-to-modern-language migration benchmarks, GPT-6.1 Sol achieves 86.6 percent accuracy at low reasoning effort, climbing to 97.9 percent when maximum reasoning is applied.

How does legacy migration work in practice?

Engineering teams dealing with legacy codebases can now translate outdated COBOL logic into safer languages like Rust, Java, or C++ with unprecedented precision. Instead of manual line-by-line rewrites that take months, developers can use these models to handle structural translation while focusing human oversight purely on security boundaries and edge-case validation.

Is it worth adopting for your codebase?

For financial institutions and enterprises bound to legacy systems, these numbers eliminate the primary risk of AI-assisted refactoring: silent syntax errors in critical financial logic. However, teams must still budget for extensive integration testing, as high migration accuracy does not automatically eliminate underlying architectural flaws in decades-old software.