From OpenAI · · 1m 10s
OpenAI advances coding in low-resource languages
OpenAI's new models improve COBOL, Fortran and Visual Basic coding with higher accuracy and lower costs.
In brief
OpenAI's new models improve COBOL, Fortran and Visual Basic coding with higher accuracy and lower costs. OpenAI's latest models achieve 86.6% accuracy in migrating COBOL code while costing 39% less than previous versions. Originally reported by OpenAI.
Low-resource languages critical for infrastructure
COBOL supports 40% of online banking systems, 80% of credit card transactions and 95% of atm transactions globally.
AI struggles with legacy codebases
Developers report coding models often struggle with low-resource languages like cobol and fortran.
MultiLangBench evaluation
GPT-6 Astra achieves 87.4% test pass rate across 20 languages, up from 26.8% for GPT-5.6 Sol.
COBOL migration results
GPT-6.1 Sol achieves 86.6% accuracy at low reasoning effort, 39% cheaper than GPT-5.6 Sol.
Cost efficiency improvements
GPT-6 Luna reaches 40.3% accuracy at 78% lower estimated api cost than GPT-5.6 Sol.
Behavior preservation in migration
Migrating COBOL programs requires preserving business logic and language-specific behaviors like rounding rules.
OpenAI releases evaluation tools
OpenAI shares MultiLangBench test suite and results for developers to inspect and identify gaps.
Next steps for enterprise software
Teams can use OpenAI's API to try coding tasks in their preferred languages while providing feedback.
The takeaway
OpenAI's latest models achieve 86.6% accuracy in migrating COBOL code while costing 39% less than previous versions.
