OpenAI and Anthropic Push AI Tools for COBOL Migration, but Experts Urge Caution

OpenAI announced in October 2026 that its latest AI models have made “substantial improvements” in migrating legacy code written in older programming languages into modern alternatives — a development that could affect critical infrastructure worldwide still running on decades-old software.

COBOL, a programming language created in 1959, remains embedded in vast portions of global infrastructure. According to IBM, it underpins more than 40 percent of online banking systems, 80 percent of in-person credit card transactions, and 95 percent of ATM transactions. It also supports U.S. government systems including Social Security and the IRS. One estimate puts the annual cost of COBOL’s inefficiencies to the U.S. GDP at around $105 billion.

In its own tests, OpenAI’s GPT-6.1 Sol model achieved 87 percent behavior preservation accuracy when migrating COBOL to modern languages such as Python, Java, and Rust at low reasoning effort, rising to 98 percent at higher reasoning. Earlier models performed significantly worse, with GPT-5.6 Sol reaching only 5 percent at low reasoning effort. OpenAI spokesperson Oscar Haines acknowledged the problem “is still far from a solved problem.”

Anthropic also entered the space in February 2026, publishing a COBOL modernization guide and pitching its Claude Code agent for legacy migrations, though it offered no performance benchmarks. IBM shares fell following Anthropic’s announcement.

Despite the progress, software engineering experts told The Verge that current AI capabilities remain far from sufficient for safely replacing critical legacy systems. Arie van Deursen, a software engineering professor at Delft University of Technology, called OpenAI’s blog “superficial” and said its evaluations were “lightyears” from actual modernization challenges. Vadim Zaytsev, an associate professor at the University of Twente, noted that “AI is not playing a major role at the moment in replacing or migrating legacy COBOL systems,” pointing to complexities including multi-language codebases, data formats, and system integrations.

Experts stressed that robust testing and human verification remain essential. As Zaytsev put it, even a theoretical tool that handles 99 percent of a migration still leaves a potential one million lines of problematic code — and “in many cases even minor differences in system behaviour before and after migration are unacceptable.”