Overview
- OpenAI announced model improvements on Monday and said GPT-6.1 Sol reached 97.9% accuracy on its COBOL migration evaluation when run with higher reasoning effort.
- The company reported broad gains for GPT-6 Astra on MultiLangBench, with an average pass rate rising to 87.4% and language-specific jumps such as Fortran to 83.5% and Visual Basic to 89.4%.
- OpenAI said GPT-6.1 Sol cut API costs in the migration setting by about 39% and that GPT-6 Luna offers much lower per-sample costs to let teams run more experiments affordably.
- Despite high benchmark scores, the firms and reporters caution that migrated programs can still change behavior without syntax errors so independent testing and human verification remain essential for banks, governments, and other regulated users.
- The push responds to a shrinking pool of legacy-language experts and large volumes of production COBOL and Fortran code, and it puts OpenAI in direct competition with vendors such as Anthropic and IBM for enterprise modernization work.