Details
- OpenAI says its latest models are better at building and migrating software written in low-resource languages.
- The announcement covers COBOL, Fortran and Visual Basic, which underpin systems used by banks, governments and businesses.
- OpenAI describes these languages as low-resource because models have less publicly available code and documentation to learn from than Python or JavaScript.
- The company reports a new MultiLangBench evaluation covering 20 languages and database-engine implementation tasks.
- On that benchmark, GPT-6 Astra achieved an 87.4% average test pass rate, compared with 26.8% for GPT-5.6 Sol.
- OpenAI also evaluated COBOL migration into newer languages while requiring the converted software to preserve the original behavior.
- Reported COBOL migration accuracy reached 86.6% at lower reasoning effort and 97.9% at higher effort, versus 5.4% for GPT-5.6 Sol at lower effort.
- The results could help organizations analyze or modernize legacy systems, although production migrations still require testing and expert review.
Impact
The announcement targets a costly modernization bottleneck: critical systems remain dependent on languages with limited documentation and shrinking specialist talent. Higher migration accuracy could reduce the effort needed to move legacy workloads into newer languages, widening adoption of AI-assisted maintenance. However, benchmark gains do not eliminate risks involving undocumented business rules, numerical behavior or security, so banks and governments are likely to use these models as supervised engineering tools rather than autonomous replacement systems.