OpenAI report: GPT-6 Astra scores 87.4% in low-resource coding test

COBOL and Fortran are old programming languages with little public material, yet they still power banks, ATMs, weather forecasts, and nuclear safety systems. Keeping them running and updated has long been a tough problem.

OpenAI released a research report on how its GPT-6 Astra model handles these languages. It also launched tests like MultiLangBench, asking models to build a SQLite-compatible database engine in 20 languages, or move COBOL code to Java, Python, and more.

The results show GPT-6 Astra reached an average pass rate of 87.4%, way up from GPT-5.6 Sol's 26.8%. COBOL jumped from 10.9% to 72.1%. In COBOL migration tests, GPT-6.1 Sol hit 97.9% accuracy at high reasoning effort.

OpenAI says this helps companies maintain old systems and modernize software, but AI migration results still need expert review and testing before real-world use.