OpenAI Enhances Code Writing and Migration Skills in Legacy Languages Like COBOL and Fortran

Summary

OpenAI announced that its latest model, GPT-6 Luna, significantly outperforms GPT-5.6 Sol in writing and porting code in legacy, low-resource languages such as COBOL, Fortran, and Visual Basic. This claim was validated through two internal benchmarks created by the company: MultiLangBench and a COBOL porting evaluation.

Why it matters

Decades-old COBOL and Fortran systems still form the backbone of banking, public sector, and scientific infrastructure worldwide. Modernizing these systems is both costly and risky. If large language models (LLMs) can reliably generate and port code in these data-scarce languages, the economic barrier to modernizing critical legacy systems would be significantly lowered. This could reshape the structure of the multi-billion-dollar IT maintenance market.

Highlights

• Two new evaluations were introduced: MultiLangBench (building a SQLite-compatible database engine in 20 languages) and a test for porting from COBOL to other languages.
• According to OpenAI's own benchmarks, GPT-6 Luna outperforms GPT-5.6 Sol in all tested languages at a lower cost.
• The MultiLangBench test package and configurations have been published on GitHub for external auditing.

Critical take

Both benchmarks were designed and executed by OpenAI. Therefore, the claim of 'significant improvement' relies on the company's own selected test cases rather than independent, industry-standard evaluations. A SQLite-compatible database engine is vastly different from the complex, undocumented COBOL mainframe code running in production environments. Consequently, the applicability of these results to the real world remains uncertain.