OpenAI Models Jump Sharply on COBOL, Fortran and Visual Basic Coding
OpenAI reports that its newest models write code in rare and outdated programming languages far better than earlier generations, on languages that still run banking systems, ATMs and critical infrastructure. On OpenAI’s benchmark, GPT-6 Astra passes 87.4% of tests against 26.8% for GPT-5.6 Sol, with the gains concentrated in the least common languages:
- COBOL: 10.9% to 72.1%
- Fortran: 21.4% to 83.5%
- Visual Basic: 27.5% to 89.4%
GPT-6.1 Sol reaches 97.9% accuracy when porting programs from COBOL to modern languages. To measure this, OpenAI built MultiLangBench, in which a model must implement a SQLite equivalent in one of 20 programming languages; the benchmark is open-sourced.
The commercial case is that large volumes of financial and government code were written decades ago, maintenance is expensive, and the engineers who can read it are retiring. The models are approaching the point of not only understanding such code but moving it to Python, Java or Rust while preserving the original logic.