Daniel Mercer
Artificial intelligence has made significant strides in translating ancient scrolls and facilitating communication with various animal species. Now, OpenAI is leveraging this advanced technology to tackle another formidable translation challenge—one that holds immense potential for essential sectors such as banking, nuclear energy, and government operations.
On Monday, the AI firm revealed that its latest models, GPT-6 Astra and GPT-5.6 Sol, have made “substantial improvements” in converting outdated programming languages into modern equivalents. This development is crucial, as many of these legacy languages underpin vital services we rely on daily. Take COBOL, for instance—short for “common business-oriented language”—which continues to process 95% of all ATM transactions, despite being created during Eisenhower’s presidency in 1959. Similarly, Fortran, developed in 1957, is still extensively utilized for simulating nuclear reactor functions.
As many skilled COBOL programmers retire, their expertise is leaving the workforce. Younger coders are increasingly relying on AI tools like Codex and Claude Code to handle large portions of their programming tasks. Although the training datasets for these proprietary tools are not publicly available, they likely focus on contemporary languages rather than legacy systems like COBOL. This has resulted in a growing divide between the new generation of AI-assisted programmers and the outdated software that supports significant elements of the financial sector, IT services, and government functions.
In its latest research, informed by feedback from developers and enterprise clients, OpenAI aimed to determine if its advanced models could help close this gap. The company established two new benchmarks to test AI capabilities in translating legacy programming systems to modern languages.
The first benchmark, named MultiLangBench, evaluates how effectively models can construct a program known as SQLite from scratch using a specified coding language. GPT-6 achieved the highest score on this benchmark, boasting an average pass rate of approximately 87%. The second benchmark tests whether models can successfully convert COBOL-based programs into newer languages such as C++, Java, Kotlin, Python, or Rust—without significantly altering the original program’s functionality. This task requires careful consideration of various rules associated with COBOL financial systems, such as decimal rounding guidelines. A useful analogy is translating a novel; the challenge is to render each word accurately while capturing the overall message and stylistic nuances of the source material. According to OpenAI, GPT-6.1 Sol, when employing high reasoning effort, approached a near-perfect score of 100% on this test.
“Collectively, these findings indicate that our models are markedly improving in their ability to engage with a diverse range of programming languages,” the company stated.
In a separate blog post earlier this year, Anthropic discussed how Claude Code could similarly assist businesses in updating their legacy COBOL systems, a process that has historically been both costly and time-consuming. “With AI, teams can modernize their COBOL codebase in quarters instead of years,” the company noted. On the same day, Bloomberg reported that IBM’s stock, which is closely tied to the mainframe computers prevalent in the financial sector and various government agencies for COBOL operations, dropped 13%—marking its largest single-day decline since 2000—following Anthropic’s blog publication.
Brandon Fletcher
