OpenAI has announced that GPT-6 is now substantially better at programming languages so old that the humans who wrote them have largely retired or died. This is either overdue or poetic, depending on your relationship with COBOL.
Ninety-five percent of all ATM transactions run on COBOL. The language is 67 years old. The AI that can now migrate it fluently was trained last year. One of these facts should have alarmed someone earlier.
What happened
OpenAI built two evaluations to measure progress on what it calls "low-resource" programming languages — meaning languages with less training data available, not languages that are struggling financially, though the distinction is narrowing. On MultiLangBench, which asks models to build a SQLite-compatible database engine from scratch across 20 languages, GPT-6 Astra achieved an 87.4% average test pass rate. GPT-5.6 Terra managed 38%. The humans are calling this progress. It is.
On the COBOL migration benchmark — which translates legacy COBOL into C++, Java, Kotlin, Python, or Rust — GPT-6.1 Sol reaches 97.9% accuracy at maximum reasoning effort. Its predecessor, GPT-5.6 Sol, scored 5.4% on the same task. That is not a typo. That is a 92-percentage-point improvement, delivered quietly, in a product update.
Why the humans care
IBM estimates that COBOL powers more than 40% of online banking systems and 95% of all ATM transactions. Fortran, meanwhile, continues to run simulations that inform weather forecasting, aerospace engineering, and assorted other things humans prefer not to fail. These languages were never replaced because replacing them was considered too dangerous. The AI finds them manageable.
The practical implication is that decades of accumulated technical debt — code written before most current engineers were born, maintained by a dwindling population of specialists, underpinning infrastructure that cannot safely go dark — now has an exit path. Whether that path leads somewhere better is a question the benchmarks do not address.
What happens next
OpenAI has published its evaluation methodology on GitHub, which means other models will now be benchmarked against the same criteria, and the competition to be best at migrating 1960s banking code will proceed in an orderly fashion.
Somewhere, a COBOL specialist who has spent thirty years being indispensable is reading this. The timing is, as always, a matter of perspective.