OpenAI Says New Models Handle COBOL, Fortran Better

OpenAI says its newest models have become much better at writing and migrating code in COBOL, Fortran, Visual Basic and other “low-resource” programming languages. In a post on 19 October 2026, the company published results from two of its own evaluations. One asks models to build a database engine from scratch in 20 languages; the other asks them to migrate COBOL programs to modern languages without changing how they behave.

The scores are OpenAI’s, measured on benchmarks OpenAI designed. They point at a real problem, though. Much of the world’s banking and public infrastructure still runs on decades-old code, and the engineers who understand it are increasingly scarce. OpenAI says its work was guided by developers and enterprise customers who told it that coding models often struggle with these languages.

In brief

OpenAI says GPT-6 Astra reached an average 87.4% test pass rate across 20 languages on its MultiLangBench evaluation, up from 26.8% for GPT-5.6 Sol.
On OpenAI’s COBOL migration evaluation, it says GPT-6.1 Sol scored 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol, rising to 97.9% at higher effort.
OpenAI is releasing the MultiLangBench setup and test suite on GitHub, and says production use still requires expert review and testing.

Why low-resource languages matter

OpenAI describes low-resource languages as those with less code and documentation for models to learn from than languages such as Python and JavaScript. Yet they “support critical infrastructure and essential services around the world”, it says. Its post 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 notes that Fortran is used in weather forecasting and nuclear reactor safety analysis.

Maintaining this software is hard, the company argues: documentation may be incomplete, the original developers may have left, and “important business rules may only exist in code.” Whether teams are fixing bugs or migrating, they need to preserve the behaviour their systems depend on.

Building SQLite from scratch, by OpenAI’s numbers

MultiLangBench asks a model to build a SQLite-compatible database engine in a given language. The model must organise a codebase, manage state, handle errors and add features without breaking existing ones. OpenAI says it uses database files as a shared interface and checks results against the reference SQLite implementation, using tests adapted from SQLite’s own test suite.

At maximum reasoning effort, OpenAI says GPT-6 Astra improved in every language tested. It cites COBOL rising from 10.9% to 72.1%, OCaml from 25.6% to 85.8%, Fortran from 21.4% to 83.5% and Visual Basic from 27.5% to 89.4%. “Astra scores above 83% in 19 of the 20 languages, with COBOL remaining the hardest at 72%,” the post says.

The cheaper GPT-6 Luna reached a 40.3% average, according to OpenAI, at an estimated average API cost of $0.77 per sample. That compares with $3.48 for GPT-5.6 Sol, which the company says is about 78% lower.

Language	GPT-5.6 Sol	GPT-6 Astra
COBOL	10.9%	72.1%
OCaml	25.6%	85.8%
Fortran	21.4%	83.5%
Visual Basic	27.5%	89.4%
20-language average	26.8%	87.4%

MultiLangBench test pass rates at maximum reasoning effort, as reported by OpenAI.

The COBOL migration test

The second evaluation asks models to migrate COBOL programs to C++, Java, Kotlin, Python or Rust. OpenAI notes that small language details carry business meaning. COBOL’s fixed-point arithmetic, for example, discards excess decimal digits unless rounding is explicitly requested, so a rewrite that rounds can run cleanly yet produce different financial results. Hidden tests require “byte-for-byte parity on observable outputs”, and a migration counts as a success only if every test passes.

On this test, OpenAI says GPT-6.1 Sol scored 86.6% at low reasoning effort, against 5.4% for GPT-5.6 Sol at the same setting, “at just 61% of the API cost”. At higher reasoning effort, it reports accuracy of 97.9%, bringing its own evaluation close to saturation. The company adds a clear limit: applying these capabilities to production software “still requires expert review and testing against the system’s business rules and dependencies.”

Our read

Publishing the MultiLangBench test suite on GitHub is the most useful part of this announcement, because it lets others check OpenAI’s numbers rather than take them on trust. The jumps are striking, but they come from the vendor’s own benchmarks, and the migration test has not yet been released or independently reproduced in the sources read. For banks and public bodies weighing COBOL modernisation, the realistic takeaway is faster first drafts, still checked line by line by people who know the business rules.