OpenAI Says Its Newest AI Models Can Translate Old Business Software Into Modern Code

OpenAI said on Monday, October 26, 2026, that its latest models are much better at moving old software into modern programming languages. The company said the work is useful for "maintaining and modernizing older systems."
According to The Verge, OpenAI reported results for converting COBOL code into C++, Java, Kotlin, Python, and Rust. COBOL is an older programming language that still runs on many large business computers. The goal of each conversion is to keep the software doing exactly what it did before.
OpenAI counts a conversion as successful only if all of the tests pass. Its main measure is called behavior preservation accuracy. It shows how often the new code behaves the same as the original.
The scores vary sharply. They depend on the model and on the "reasoning effort" setting, which controls how much extra processing the model spends on a problem.
GPT-6.1 Sol reached 87 percent accuracy at low reasoning effort and 98 percent at higher effort. At low effort, GPT-6 Astra scored around 86 percent, GPT-6 Sol scored 21 percent, and GPT-5.6 Sol scored 5 percent.
More effort usually means more computing, and more computing costs more money. For a company choosing a model, the gap between effort settings can matter as much as the gap between model versions.
OpenAI had described the work earlier. On October 19, 2026, it published a post titled "Advancing coding in low-resource programming languages" on its own website. The post says its COBOL test checks whether a model can move a program to another language while keeping its behavior the same. Engadget reported on October 20, 2026 that OpenAI said its models are now better at working with COBOL and other old languages.
Anthropic, the company behind the Claude AI models, is pursuing the same kind of work. Its Code modernization page says Claude Code helps large companies update old codebases while keeping their business logic intact. Its Code Modernization Playbook describes moving from COBOL to languages such as Java or Python. A webinar titled Modernizing legacy code with Claude Code covers a COBOL-to-Java migration and an upgrade of about 500,000 lines of Java 8 code to Java 17. A separate financial services webinar lists COBOL-to-cloud migrations and mainframe modernization among its legacy topics.
The money involved became clear earlier this year. On February 24, 2026, Reuters reported that IBM shares posted their steepest daily drop since 2000. The fall followed Anthropic's statement that Claude Code could be used to modernize COBOL running on IBM systems. Reuters linked the share move to that announcement.
The broader context here is that COBOL still runs in production at many large institutions. Moving it has long been a question of cost, risk, and scarce skills, not just translation. A system that passes a set of tests has shown it works the same way only in the places those tests check. It is a bit like proofreading a translated contract: a clean read tells you about the clauses you actually checked. OpenAI's rule that all tests must pass puts the weight on the quality of those tests, which the buyer must supply or verify. The gap between reasoning settings also suggests that the headline numbers depend heavily on how a model is configured.
In my view, the more useful number for people doing the work is that gap between effort settings. If accuracy rises from 87 to 98 percent as effort increases, the practical question becomes how much extra computing a given migration is worth. Whether a checking process can catch the remaining errors before the switch-over is a problem engineers can work on. A general claim of "modernization" is much harder to act on.
The practical upside, if the reported figures hold up in real projects, is that large, well-tested old systems could become cheaper to move. Teams that have put off mainframe work for a decade may gain a realistic option. Passing tests is still not the same as a safe switch-over, and that remaining work falls to the engineers.


