OpenAI Says Its Newest Models Can Migrate Old COBOL Code
OpenAI said on Monday, October 26, 2026, that its latest models show "substantial improvements" in migrating aging software written in older programming languages into modern code. The company framed the capability as useful for "maintaining and modernizing older systems."
According to The Verge, OpenAI reported results for migrating COBOL, the business language still common on mainframes, to C++, Java, Kotlin, Python, and Rust without changing what the original software does. OpenAI stated that a migration counts as successful only if all tests pass. The metric in the company's evaluation is behavior preservation accuracy, meaning the share of migrated programs that keep the original behavior.
The reported scores vary sharply by model and by reasoning effort, the amount of computation a model spends working through a problem before it answers. GPT-6.1 Sol reached 87 percent behavior preservation accuracy at low reasoning effort and 98 percent at higher reasoning effort. At low reasoning effort, GPT-6 Astra scored around 86 percent, GPT-6 Sol scored 21 percent, and GPT-5.6 Sol scored 5 percent. For teams weighing model choice against inference cost, the gap between low and high reasoning effort on the same model is as informative as the gap between generations.
OpenAI had described the work earlier. On October 19, 2026, the company published a post titled "Advancing coding in low-resource programming languages" on its own domain. The post states that its COBOL migration evaluation tests whether a model can migrate existing programs to another language while preserving their behavior. Engadget reported on October 20, 2026 that OpenAI said its models are now better at working with COBOL and other old languages.
Anthropic has pursued the same workload. Its Code modernization page says Claude Code helps leading enterprises modernize legacy codebases while maintaining business logic integrity. Its Code Modernization Playbook describes migration from COBOL to modern 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 roughly 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 commercial stakes of this category became visible earlier this year. On February 24, 2026, Reuters reported that IBM posted its steepest daily drop since 2000 after Anthropic said Claude Code could be used to modernize COBOL running on IBM systems. Reuters tied the share move to that announcement, which put the market's attention on mainframe-hosted COBOL estates as a target for AI-assisted tooling.
The broader context here is that COBOL remains in production at many large institutions, and migration has long been a matter of cost, risk, and scarce skills rather than translation alone. A system that passes a test suite has been shown equivalent only within the coverage of those tests. OpenAI's all-tests-must-pass criterion puts the weight on the quality of the test corpus, which the enterprise buyer must supply or validate. The spread between reasoning settings also suggests that headline numbers depend heavily on configuration.
In my view, the more useful reading for practitioners is the reasoning-effort curve. If behavior preservation rises from 87 to 98 percent as reasoning effort increases, the operational question becomes how much extra compute a given migration justifies, and whether a verification harness can catch the residual failures before cutover. That is a tractable engineering problem in a way that a general claim of "modernization" is not.
If the reported figures hold up in production, large, well-tested legacy systems may become cheaper to port, and teams that have deferred mainframe work for a decade gain a realistic option. The distance between a passing test suite and a safe cutover is still work that engineers will have to do.
