OpenAI announced on October 19, 2026 that its latest models show improved performance writing and migrating software in what it calls low-resource programming languages, including COBOL, Fortran and Visual Basic. The company notes these languages have less publicly available code and documentation than Python or JavaScript, yet underpin critical systems. It cites figures that COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions, and that Fortran is used in weather forecasting and nuclear reactor safety analysis.
OpenAI developed two evaluations. MultiLangBench tests whether a model can build a SQLite-compatible database engine across 20 languages, using database files as a shared interface to apply the same tests regardless of language. A separate COBOL migration evaluation tests whether a model can convert programs to C++, Java, Kotlin, Python or Rust while preserving behavior, using hidden tests requiring byte-for-byte parity on outputs.
According to OpenAI, GPT-6 Astra reaches an average 87.4% test pass rate on MultiLangBench, up from 26.8% for GPT-5.6 Sol. On the migration evaluation, GPT-6.1 Sol reports 86.6% accuracy at low reasoning effort and up to 97.9% at higher effort. OpenAI says it is releasing the MultiLangBench setup and test suite on GitHub. All figures are reported by OpenAI and are not independently verified.
- GPT-6 Astra reports an average 87.4% MultiLangBench pass rate across 20 languages, up from 26.8% for GPT-5.6 Sol (OpenAI's own figures)
- On COBOL migration, GPT-6.1 Sol reports 86.6% accuracy at low reasoning effort, up to 97.9% at higher effort
- OpenAI cites that COBOL supports over 40% of online banking systems and 95% of ATM transactions
- MultiLangBench evaluation setup and test suite are being released on GitHub
What it means for you
If your organization still runs on COBOL, Fortran or Visual Basic — common in banking, insurance, government and older manufacturing systems — this says AI tools are getting better at reading, fixing and rewriting that code. The important detail buried here is the word 'behavior-preserving': migrating legacy code isn't about making it run, it's about making it produce the exact same results, down to rounding rules on financial transactions. These are vendor benchmarks, not a proven track record on your specific codebase, so treat the numbers as a direction of travel, not a guarantee.
Try this
If you have legacy systems nobody fully understands anymore, spend an hour writing down what you actually know about one of them: what it does, who maintained it, where the business rules live, and what breaks if it stops. That map is worth more than any migration tool, and you'll need it either way.
Who should care
Owners and IT leads at organizations running mainframe or legacy code — banks, insurers, utilities, government, older manufacturers — and anyone weighing a modernization or migration project.
Skip this if
Your systems run on modern, mainstream languages, or you have no legacy code and no migration on the horizon.
Sources: OpenAI — read the original