OpenAI says GPT-6 improves COBOL and Fortran performance
OpenAI says its latest models are substantially better at understanding and translating low-resource languages such as COBOL and Fortran.
OpenAI says it has improved its latest models at working with low-resource programming languages, including COBOL, Fortran and Visual Basic. These languages have far less publicly available training code than Python, but they still support important systems, according to the company’s announcement.1
The announcement frames the work around two evaluations. MultiLangBench asks models to build a SQLite-compatible database engine from scratch across 20 programming languages. OpenAI reports that GPT-6 Astra reached an average pass rate of 87.4% at maximum reasoning effort, compared with about 38% for GPT-5.6 Terra.1
OpenAI also describes a COBOL migration benchmark. It tests whether a model can translate COBOL into C++, Java, Kotlin, Python or Rust. The company reports that GPT-6.1 Sol reached 86.6% accuracy at low reasoning effort, rising to 97.9% when given more reasoning time, compared with 5.4% for GPT-5.6 Sol.1
The wider point is less glamorous than a new chatbot demo, but arguably more useful. Replacing or maintaining old code is a major problem for banks, public-sector systems and other organisations that cannot simply discard decades of working infrastructure. The Decoder reports the same figures while highlighting the practical challenge of modernising code written in languages that have become specialist skills.2
That does not mean the benchmarks prove that an AI system can safely rewrite a live banking platform. Translation accuracy is only one part of a migration: teams still need tests, human review, security checks and a way to preserve undocumented behaviour. The results are better understood as evidence that models may be becoming more useful for the first-pass analysis and conversion work around legacy systems, rather than as permission to remove engineers from the process.
OpenAI’s announcement also makes the cost trade-off visible. The reported gains vary by model and reasoning setting, so the most capable result is not necessarily the cheapest or fastest option for every migration. Organisations will need to compare accuracy with review costs, failure risk and the difficulty of verifying the output in the target language.
Our opinion
This is the kind of AI progress that deserves more attention than another synthetic image demo. COBOL and Fortran are not fashionable, but they remain embedded in systems people rely on. Better assistance with those languages could help teams understand ageing code and reduce the drudgery of migration. The sensible conclusion is not that legacy developers are obsolete; it is that they may finally have a more capable tool for the least rewarding parts of the job.
The catch is that OpenAI’s numbers are company-reported benchmark results, not a guarantee of production safety. Until the tests, code and failure cases are independently examined, GPT-6’s legacy-language performance should be treated as promising evidence rather than a finished migration solution.