Pull cash from an ATM and the transaction quite likely runs on a programming language created in 1959, while Eisenhower was still president. COBOL (the name stands for “common business-oriented language”) processes 95% of all ATM transactions to this day. OpenAI now claims its newest models can convert that sort of code into something today’s developers would recognize.
On Monday, the company said its latest models, GPT-6 Astra and GPT-5.6 Sol among them, have made “substantial improvements” at moving old, outdated programming languages over to newer ones. The pitch reaches past banking into other critical infrastructure, including nuclear energy and government.
A near-perfect score, with conditions
OpenAI’s COBOL test produced the headline figure. According to OpenAI, GPT-6.1 Sol running with high reasoning effort scored very close to 100%.
The test asks a model to take programs run on COBOL and rewrite them in C++, Java, Kotlin, Python or Rust while keeping the program’s behavior essentially the same. It’s tougher than it looks. COBOL-based financial systems come with their own rules, like how and when decimals are rounded, so a model needs to understand that wider context instead of simply trading one syntax for another.
A rough comparison is translating a novel: every word has to be right, and the original’s meaning and tone have to survive too. If a banking system loses a rounding rule, the mistake ends up in customers’ account balances.
There’s a catch. OpenAI designed this benchmark and reported the results itself. Acing a vendor’s in-house test is not the same as proving the model works inside a bank’s mainframe.
Two new benchmarks aimed at the same problem
The research was “guided by feedback from developers and enterprise customers,” OpenAI said. It led to two new benchmarks, each meant to gauge how well AI can carry legacy systems from one language into another.
MultiLangBench, the first of them, tests how well models can build a program called SQLite from scratch in a given coding language. GPT-6 came out on top, with an average pass rate of about 87%.
“Together, these results show that our models are becoming significantly better at working with a wide range of programming languages,” the company wrote.
Old code, fewer people who know it
COBOL has company among the antiques still in use. Fortran dates to 1957 and is still widely used to simulate how nuclear reactors work.
The experts in these languages are disappearing. Plenty of COBOL-trained software engineers have retired and taken their knowledge with them. Younger programmers, meanwhile, pass big portions of their work to AI assistants such as Codex and Claude Code.
Nobody outside these companies knows exactly what data the proprietary tools were trained on. Still, they were very likely trained mostly on modern languages and saw far less legacy code like COBOL. The result is a dangerous gap, and a growing one, between a new wave of AI-assisted coders and the software that runs much of the financial sector, IT services and government agencies.
Anthropic made the pitch first, and IBM paid for it
OpenAI has competition here. Earlier this year, Anthropic said in a blog post that Claude Code could help businesses update aging COBOL systems, work that has traditionally been slow and costly.
“With AI, teams can modernize their COBOL codebase in quarters instead of years,” Anthropic wrote.
Investors reacted. Bloomberg reported that IBM stock dropped 13% on the day Anthropic published the post, its largest single-day fall since 2000. IBM mainframes are widely used to run COBOL in the financial sector and at many government agencies.
That puts the two leading AI labs after the same legacy code. The near-100% result comes from OpenAI’s own test, and a bank looking to replace a system older than its CTO should run its own checks before relying on it.