OpenAI Lifts Low-Resource Coding Accuracy to 87.4% Across 20 Languages, Cutting Costs
Updated
Updated · OpenAI · Oct 19
OpenAI Lifts Low-Resource Coding Accuracy to 87.4% Across 20 Languages, Cutting Costs
3 articles · Updated · OpenAI · Oct 19
Summary
GPT-6 Astra reached an 87.4% average pass rate on MultiLangBench across 20 languages, up from 26.8% for GPT-5.6 Sol, while GPT-6 Luna hit 40.3% at about 78% lower estimated API cost.
The gains were strongest in low-resource languages used in critical systems: COBOL rose to 72.1% from 10.9%, Fortran to 83.5% from 21.4%, and Visual Basic to 89.4% from 27.5%.
On COBOL migration, GPT-6.1 Sol achieved 86.6% accuracy at low reasoning effort versus 5.4% for GPT-5.6 Sol, and reached 97.9% at higher effort while costing 39% less.
OpenAI said the work targets software modernization where documentation is sparse and business rules live in legacy code; COBOL still underpins more than 40% of online banking, 80% of in-person card transactions and 95% of ATM transactions.
The company is releasing the MultiLangBench setup and test suite on GitHub, while stressing production migrations still require expert review and testing against business rules and dependencies.