OpenAI Says Its Models Can Give the Financial Sector a Coding Makeover

Artificial intelligence has deciphered ancient scrolls and could even help humans communicate with other animals, but now OpenAI is using the technology for another difficult translation task—one that could have huge benefits for critical infrastructure like banking, nuclear energy, and government.

The AI startup announced Monday that its latest models, including GPT-6 Astra and GPT-5.6 Sol, have shown “substantial improvements” in their ability to migrate old, outdated programming languages into newer ones. That’s potentially a big deal, since some of those dated languages are the lifeblood of some basic services we all depend on. COBOL, for example—short for “common business-oriented language”—still handles 95% of all ATM transactions, even though it was invented when Eisenhower was in the Oval Office (1959). Fortran, a programming language invented in 1957, is still widely used to simulate the functionality of nuclear reactors.

While many software engineers trained to use COBOL have been retiring—and bringing their skills with them—younger programmers have been delegating large portions of their work to AI assistants like Codex and Claude Code. While the datasets of those proprietary tools aren’t public, they were very likely trained using more contemporary programming languages, rather than with “legacy” languages like COBOL. The result has been a dangerous and ever-widening gap between the a new generation of AI-assisted coders and the software running a considerable portion of the financial sector, IT services, and government agencies.

OpenAI tests new benchmarks

In new research “guided by feedback from developers and enterprise customers,” OpenAI set out to see if its newer models could help to bridge that gap. The company created two new model testing benchmarks, both of which are essentially intended to test AI’s ability to translate legacy systems from one programming language to another.

The first, called MultiLangBench, gauges how well models can build a program called SQLite from scratch using a particular coding language. GPT-6 scored the highest on this benchmark, with an average pass rate of about 87%. The second test assesses whether models can migrate COBOL-run programs into newer programming languages—C++, Java, Kotlin, Python, or Rust—without meaningfully altering the program’s behavior. This is a delicate dance. Financial systems using COBOL come with varying rules—such as how and when decimals should be rounded—which means OpenAI’s models needed to pick up on that broader context. In a very loose analogy, you can think of it like translating a novel: the challenge lies in translating every word as accurately as possible, while simultaneously remaining true to the gestalt message and stylistic overtones of the original text. GPT-6.1 Sol with high reasoning effort scored very close to 100% on that test, according to OpenAI.

“Together, these results show that our models are becoming significantly better at working with a wide range of programming languages,” the company wrote.

Industry impact on legacy software

Anthropic wrote in its own blog post early this year that Claude Code could similarly help businesses update outdated systems built on COBOL, a historically expensive and time-consuming process. “With AI, teams can modernize their COBOL codebase in quarters instead of years,” the company wrote. Bloomberg reported the same day that shares of IBM, whose mainframe computers are widely used throughout the financial sector and across many government agencies to run COBOL, fell 13% after Anthropic published its blog post—its biggest single-day slide since 2000.