According to a report by The Decoder, OpenAI has announced that its GPT-6 models feature significant performance upgrades for low-resource and legacy programming languages, specifically highlighting COBOL, Fortran, and Visual Basic. Performance metrics cited in the announcement show that a model named GPT-6 Astra achieved an 87.4 percent test pass rate across 20 languages on a benchmark called MultiLangBench. This compares to approximately 38 percent for the previous GPT-5.6 Terra model. Furthermore, a variant designated as GPT-6.1 Sol attained up to 97.9 percent accuracy specifically for COBOL migration tasks.
Significance for Legacy Financial Infrastructure
Legacy programming languages like COBOL and Fortran remain deeply embedded in the foundational software of critical sectors, particularly banking, finance, and government operations. Much of this technology dates back decades, running on aging mainframes while facing a shrinking pool of human programmers who understand the codebases. The ability of modern language models to accurately read, translate, and migrate legacy code presents a potential pathway for organizations seeking to update critical systems without undergoing entirely manual rewrites. Higher accuracy rates reduce the risk of translation errors that could cause severe system failures in production environments.
Unresolved Details and What Comes Next
Despite the released accuracy figures and benchmark scores, the announcement left several practical questions unanswered. The company did not specify the exact timeline for public availability of GPT-6 Astra and GPT-6.1 Sol, nor did it disclose pricing structures or API availability details for enterprise clients. Observers and IT leaders will need to watch for real-world deployment case studies to determine whether these laboratory benchmark scores translate into reliable performance during large-scale enterprise codebase migrations.
As enterprises continue to evaluate generative artificial intelligence for software engineering tasks, the practical utility of these models on older systems will depend heavily on independent verification of the reported migration accuracy.
Lead image: Brett Sayles / Pexels
