OpenAI\'s GPT-6 series showed much stronger coding performance than earlier models in programming languages with limited training data, including COBOL, Fortran and Visual Basic. OpenAI released results from a new MultiLangBench benchmark that measures how well models build an SQLite-compatible database engine across 20 languages and pass set functional tests. OpenAI said the results could support maintenance and modernization of legacy software, while noting the evaluation was internally designed and would require expert review and further validation for real systems.