COBOL migration turned a corner in "Advancing coding in low-resource programming languages," the official OpenAI research post published on October 19, 2026. The post reports results from two new evaluations: GPT-6.1 Sol migrated COBOL programs to modern languages with 86.6% accuracy at low reasoning effort and 97.9% at higher effort, while the previous GPT-5.6 Sol scored just 5.4% at the same low setting. For the banks and government agencies still running decades-old systems, a jump of that size matters more than another leaderboard gain in a mainstream language.
The tests measure real work, not memorized syntax
OpenAI built two proving grounds. The first, MultiLangBench, asks a model to build a SQLite-compatible database engine from scratch, implemented separately in 20 different programming languages, testing the long-horizon engineering skills of organizing code, managing state, handling errors, and adding features without breaking existing functionality. The second is a COBOL migration evaluation: existing COBOL programs are migrated to C++, Java, Kotlin, Python, or Rust, and hidden tests require byte-for-byte parity on observable outputs — every test must pass, because a small change in business logic means the books no longer balance. One benchmark tests building from zero, the other tests faithful preservation, which are exactly the two jobs legacy teams do most.
The biggest gains are in the languages with the least training data
On MultiLangBench, GPT-6 Astra reaches an average test pass rate of 87.4% across the 20 languages, up from 26.8% for GPT-5.6 Sol, with improvements in every language tested. The per-language view is starker: COBOL rises from 10.9% to 72.1%, Fortran from 21.4% to 83.5%, Visual Basic from 27.5% to 89.4%, and OCaml from 25.6% to 85.8%. Astra scores above 83% in 19 of the 20 languages, with COBOL remaining the hardest at 72.1%. The cheaper option improved too: GPT-6 Luna averages 40.3%, ahead of the previous flagship, at an estimated API cost of about $0.77 per sample versus $3.48 for GPT-5.6 Sol — roughly 78% lower. On the migration evaluation, GPT-6.1 Sol's 86.6% also costs only 61% of what the previous generation cost.
Why finance and public systems carry the stakes
Low-resource languages are not short on relevance; they are short on room for error. The figures OpenAI cites are that COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions, while Fortran is used in weather forecasting and nuclear reactor safety analysis. The maintenance problem is structural: documentation is incomplete, the original developers have left, and critical business rules exist only in the code. The example in the post is telling: COBOL fixed-point arithmetic discards excess decimal digits unless rounding is explicitly requested, so a migration tool that quietly switches to rounding produces a program that runs without errors but computes different money. The byte-for-byte requirement exists to catch exactly that kind of silent divergence.
Can this migrate a production system today?
Not on its own. OpenAI states the boundary clearly in the post: applying these capabilities to production software still requires expert review and testing against the system's business rules and dependencies. The scores also come from OpenAI's own evaluations; the external check available is to rerun MultiLangBench, whose evaluation setup and test suite have been released on GitHub together with per-language results. Across the industry, coding-model competition is leaving the comfort zone of mainstream languages: DeepSeek-V4.1-Flash previously reached 77 on DeepSWE, another benchmark built around realistic engineering tasks. The prudent path for legacy teams is a small pilot: pick one module with test coverage, migrate it, prove behavioral parity with their own tests, and only then widen the scope.
