COBOL has outlived nearly every generation of computer hardware around it, and OpenAI now believes artificial intelligence may help organizations move some of that critical software into modern languages. In a new evaluation published October 19, the company said its latest coding models made a major leap at preserving the behavior of aging COBOL programs during migration.
The headline result comes from GPT-6.1 Sol, which OpenAI says reached 86.6% accuracy at low reasoning effort on its COBOL migration test, compared with 5.4% for the previous GPT-5.6 Sol model. Increasing the reasoning setting pushed the newer model to 97.9%, while estimated API costs were 39% lower than the earlier model’s. The test required translations into C++, Java, Kotlin, Python, or Rust—not merely code that compiles, but code that produces exactly matching observable results.
That distinction matters because legacy software often contains undocumented business rules that are easy to break during a rewrite. OpenAI specifically points to COBOL’s fixed-point arithmetic: excess decimal digits can be discarded unless rounding is explicitly requested, meaning a seemingly sensible modernization could silently change a financial calculation. The company’s hidden tests therefore demand byte-for-byte parity, including details such as rounding and the handling of missing records.
The potential payoff is enormous. OpenAI says COBOL still supports more than 40% of online banking systems, 80% of in-person credit-card transactions, and 95% of ATM transactions. Banks, insurers, airlines, tax agencies, and other institutions have spent decades building dependable systems around the language, but many face incomplete documentation and a shrinking pool of engineers who understand the original code. A successful AI-assisted migration could let teams preserve those systems’ rules while gradually shifting them onto platforms that are easier to maintain.
The COBOL test is only part of OpenAI’s broader push into “low-resource” programming languages. Its MultiLangBench evaluation asks models to build a SQLite-compatible database engine in 20 languages, including COBOL, Fortran, OCaml, and Visual Basic. GPT-6 Astra achieved an average test pass rate of 87.4%, up from 26.8% for GPT-5.6 Sol; in COBOL specifically, the score rose from 10.9% to 72.1%. OpenAI has released the benchmark setup and test suite on GitHub so developers can inspect the methodology and compare results.
Those numbers are promising, but they are not permission to let a chatbot rewrite a bank’s core systems unattended. OpenAI acknowledges that production migrations still require expert review, extensive regression testing, and validation against dependencies and business requirements that may not appear in a single source file. The more realistic near-term scenario is an AI coding assistant that maps unfamiliar repositories, proposes translations, runs tests, and helps human engineers tackle the backlog—turning COBOL modernization from an all-or-nothing replacement project into a controlled, testable process.
