OpenAI said on October 19, 2026, that its latest models have improved at writing and migrating code in COBOL, Fortran, and Visual Basic, languages the company describes as low-resource because relatively little code exists for models to learn from1,2,3.
The update is a capability improvement to existing models, not a standalone product. OpenAI's technical account directed developers to try the models through its API.
OpenAI framed the work around two tasks: helping teams understand business logic written decades ago, and preserving how existing systems behave during modernization. The company is also soliciting input from developers and enterprise customers on which other languages or workflows could benefit from similar treatment.
An estimated 800 billion lines of COBOL remain in production, powering systems at banks, government agencies, and large enterprises. COBOL has historically been a weak spot for coding models; research benchmarks have shown Pass@1 and compilation rates on COBOL tasks ranging from single digits to about 40%.
OpenAI's results show progress on its own evaluations, but production safety and enterprise adoption remain unresolved, according to RuntimeWire's assessment. The company did not publish third-party benchmark comparisons or detail specific accuracy figures for the updated models.
ANALYSIS The absence of independent benchmarks or named enterprise deployments limits how much weight practitioners can assign to the claimed gains. Legacy-code modernization is a high-stakes domain where behavioral preservation matters more than raw code generation, and OpenAI's own framing acknowledges that gap by emphasizing behavior-preserving migration rather than greenfield coding.
Separately, OpenAI announced new features for ChatGPT Slides on the same weekend, including real-time collaboration and direct on-slide editing4.