OpenAI says its latest models can now build and translate software in COBOL, Fortran, and Visual Basic, languages where earlier models struggled. The claim comes from a post on OpenAI’s own website that surfaced on 19 October 2026. Both tests behind it were designed by OpenAI, and every score is measured against OpenAI’s own older models, not against a rival.

The first test, MultiLangBench, has a model write, in each of 20 languages, a database engine that works like SQLite, starting from nothing. GPT-6 Astra, with reasoning effort set to its maximum, passes 87.4% of the tests on average. GPT-5.6 Sol, the older model, managed 26.8%. A cheaper model, GPT-6 Luna, reaches 40.3% at an average of $0.77 per sample against $3.48 for GPT-5.6 Sol, which OpenAI calls roughly 78% lower. The cost is OpenAI’s own estimate.

The second test covers migration. Models must move COBOL programs into C++, Java, Kotlin, Python, or Rust, and hidden tests demand output that matches byte for byte. One failed test sinks the whole migration. GPT-6.1 Sol scores 86.6% at low reasoning effort, where GPT-5.6 Sol scored 5.4% at the same setting. OpenAI says the newer model also costs 39% less. At higher effort it reaches 97.9%, which the company itself describes as close to saturating the test.

Read those jumps for what they are. A rise from 26.8% to 87.4%, and from 5.4% to 86.6%, is a tripling or far more, on exams the vendor wrote, between its own models. A test that is nearly maxed out also has little left to separate one model from the next.

One check exists. On GitHub, OpenAI is publishing the MultiLangBench setup and its test suite, plus results for each language. The tests were adapted from SQLite’s own suite and graded against the reference SQLite. Outsiders can rerun that one. The post does not say the COBOL migration test will be released.

The stakes explain the interest. OpenAI cites outside sources for the figures. Per those sources, COBOL handles 95% of ATM transactions. It also runs online banking at more than 40% of systems, and covers 80% of credit card transactions made in person. Fortran, it adds, is used in weather forecasting and nuclear reactor safety analysis. These are third-party numbers that AI Insiders has not independently checked, not OpenAI findings.

Migration is hard for human reasons before technical ones. Teams often work with thin documentation, and the people who wrote the code may be gone. The rules a business actually runs on can live only inside the program. OpenAI gives a concrete case: COBOL’s fixed-point arithmetic drops extra decimal digits unless rounding is requested, so a rewrite that rounds will run cleanly and still produce different financial results. That is what “preserving behavior” means in practice.

COBOL also remained the weak spot in the database test. GPT-6 Astra scored 72.1% there, up from 10.9%, the lowest of its 20 languages, while it cleared 83% in the other 19. OpenAI itself says that before any production use, a specialist still has to review the output and run tests that check it against the system’s dependencies and its business rules. The company invites developers to try the models through its API.

For a bank or insurer sitting on a COBOL backlog, the sensible first step is to rerun the public benchmark and then feed the model a few of its own programs with known outputs. An 86.6% score on OpenAI’s test says little about how the model handles your rounding rules.

OpenAI, “Advancing coding in low-resource programming languages,” openai.com, surfaced 19 October 2026 (the page carries no publication date).