AI Legacy Code Migration Hits 98% Accuracy, Threatens Billion-Dollar Modernization Market

Critical infrastructure running on code older than the engineers maintaining it just got a new lifeline. OpenAI’s latest model, GPT-6 Astra, achieved a 97.9% success rate migrating COBOL applications to modern languages like Java, Python, and Rust while preserving byte-for-byte functional parity, according to benchmarks the company released this week.

The breakthrough targets a $1.5 trillion problem. Financial institutions, government agencies, and insurance giants still process trillions in daily transactions on systems written in COBOL, Fortran, and Visual Basic — languages no longer taught in computer science programs. The average COBOL developer is 55 years old. When they retire, they take institutional knowledge with them. Documentation is often missing. Contractors charge premium rates for maintenance.

“We’re not talking about generating snippets,” said an OpenAI researcher familiar with the testing. “The model understands entire system architectures built decades ago.” In the company’s COBOL migration benchmark — which requires identical observable outputs down to the byte — GPT-6.1 Sol scored 86.6% with minimal reasoning effort. With extended compute, accuracy climbed to 97.9%.

MultiLangBench, the evaluation suite OpenAI built to measure this, tasks models with building a SQLite-compatible engine across 20 languages. GPT-6 Astra averaged 87.4% success. Its predecessor, GPT-5.6 Sol, managed 26.8%. COBOL-specific scores jumped from 10.9% to 72.1%. Fortran from 21.4% to 83.5%. Visual Basic approached 90%.

The implications ripple beyond OpenAI. Legacy modernization contracts currently sustain a global consulting ecosystem worth billions annually. Firms specializing in mainframe migration, code translation, and wrap-and-renew strategies now face a tool that can automate the hardest part: understanding undocumented, decades-old logic without the original authors.

OpenAI has published MultiLangBench on GitHub for independent verification. Developers can now test the methodology against their own codebases.