OpenAI announced that its latest AI models are now more capable of handling legacy programming languages such as COBOL, Fortran and Visual Basic, a skill set that is becoming scarce as the pool of developers familiar with these languages shrinks. The company notes that roughly 40 percent of banks worldwide still depend on COBOL‑based systems, and replacing them with modern code can run into the hundreds of millions of dollars and take years to complete. With many COBOL programmers approaching retirement, the pressure to transition is increasing.
OpenAI said the improvements to its models were shaped by feedback from developers and enterprise customers who highlighted the challenges of working with older code bases. To measure progress, the firm built an evaluation framework called MultiLangBench, which asks models to generate SQLite‑compatible database engines in a chosen programming language. In the tests, the GPT‑6 Astra model achieved an average pass rate of 87.4 percent across 20 languages, a sharp rise from the 26.8 percent recorded by the earlier GPT‑5.6 Sol model. COBOL remained the toughest language, with Astra passing 72 percent of the COBOL tasks.
The company also evaluated how well its models could translate COBOL programs into modern languages such as C++, Java, Kotlin, Python or Rust while preserving the original functionality. In these migration trials, the GPT‑6.1 Sol model reached 97.9 percent accuracy when given ample reasoning time, and 86.6 percent accuracy under tighter constraints. OpenAI emphasized that a migration can only be considered successful if every test passes, because banking software must maintain exact behavior, including specific rounding rules and handling of missing records.
“Behavior‑preserving capabilities like these are useful for teams maintaining and modernizing older systems, though applying it to production software still requires expert review and testing against the system’s business rules and dependencies,” the company wrote.