OpenAI Says GPT-6 Closes Gap on Legacy Programming Languages Like COBOL and Fortran
OpenAI's latest models sharply improve performance on low-resource programming languages such as COBOL, Fortran and Visual Basic, with GPT-6 Astra reaching an 87.4 percent pass rate on the new MultiLangBench and GPT-6.1 Sol hitting up to 97.9 percent accuracy in COBOL migration.
OpenAI has substantially improved its newest models for programming languages that have long been neglected by AI training data, including COBOL, Fortran and Visual Basic. On a newly introduced benchmark called MultiLangBench, the model designated GPT-6 Astra achieves a test pass rate of 87.4 percent across 20 languages, compared with roughly 38 percent for GPT-5.6 Terra.
The gains are especially pronounced in COBOL migration, a task of considerable practical weight. According to OpenAI, the model GPT-6.1 Sol reaches up to 97.9 percent accuracy when translating or modernizing COBOL code. That result could matter well beyond the laboratory, because COBOL still underpins a large share of critical financial infrastructure, including banking and insurance systems that have accumulated decades of business logic.
Legacy languages have historically been a weak spot for large language models. Most training corpora are dominated by modern languages such as Python, JavaScript and Java, leaving older systems with comparatively little representation. COBOL and Fortran code also tends to be verbose, heavily customized and tied to specific institutional environments, which makes automated assistance harder to validate. The MultiLangBench results suggest that OpenAI has deliberately targeted this gap rather than accepting it as a permanent limitation.
The timing is significant for organizations that depend on aging systems. Banks, insurers and public administrations have long struggled to find engineers fluent in COBOL, while the programmers who built the original systems approach retirement. Migration projects are typically expensive and risky, and failures can disrupt services that customers rely on daily. A tool that can read, explain and convert legacy code with high accuracy could reduce both cost and risk, though the 97.9 percent figure also implies that a meaningful fraction of cases still require human review.
OpenAI has not disclosed how the benchmark was constructed, which languages are included in the 20-language set, or how the pass rates were measured. Independent verification will therefore be important before firms treat the numbers as a guarantee. The distinction between generating plausible code and producing correct, auditable migrations is considerable in regulated industries, where errors can carry legal and financial consequences.
Even so, the reported improvement from about 38 percent to 87.4 percent marks a step change rather than an incremental gain. If it holds up outside OpenAI's own testing, it would strengthen the case for using AI assistants in maintenance work that many technology leaders had considered too fragile for automation. For institutions shaped by decades-old software, the question is no longer whether AI can engage with COBOL and Fortran, but how carefully its output should be checked before it touches systems that keep money moving.
5
