Can AI help solve the COBOL code crisis?

COBOL dates to 1959 and still runs a remarkable share of the world. IBM says it underpins more than 40 percent of online banking systems, 80 percent of in-person card transactions and 95 percent of ATM transactions, plus government systems from Social Security to the IRS. The people who understand it are retiring or gone.

On October 19, OpenAI reported big gains in migrating legacy code into modern languages while preserving behavior. Its GPT-6.1 Sol model hit 87 percent behavior preservation at low reasoning effort and 98 percent at higher effort, far ahead of earlier models. Anthropic pitched Claude Code for the same job in February, and IBM's stock dropped after that announcement. The market clearly thinks something is shifting.

We want this to work. Legacy code is one of the most unglamorous, highest-value problems AI could help with. If models can read old systems, document them and translate them faithfully, that's real relief for banks, agencies and the taxpayers who fund decades of patchwork maintenance.

But the experts The Verge spoke with are pretty clear that today's benchmarks measure the easy part. Translating a well-defined chunk of code is not the same as untangling a system with hundreds of millions of lines, multiple languages, odd data formats and integrations nobody fully remembers. One professor called OpenAI's write-up superficial and said its tasks are lightyears away from actual modernization. The hard problem is breaking a monolith into pieces that can be swapped out safely, and the evaluations don't touch that.

The 98 percent figure is a good example of why context matters. In a consumer app, 98 percent is impressive. In a payment system, the remaining 2 percent is where people's paychecks, benefits and tax refunds live. As one expert put it, even a 1 percent remainder can mean a million lines of code failing in subtle ways. Minor behavior changes after migration are often unacceptable outright.

To its credit, OpenAI's spokesperson acknowledged legacy code is still far from solved. That honesty matters, because the real risk here isn't the technology. It's the sales pitch reaching budget-strapped agencies and banks who hear "the economics have shifted" and move faster than verification allows.

Our read: the useful near-term role for AI is as an understanding tool. Document the system, generate tests, run old and new side by side, and keep humans who know the legacy language in the review loop. Formal verification could eventually let translations be checked automatically, but that's a huge amount of work. If organizations treat AI migration as a shortcut instead of an assistant, the failures won't show up in benchmarks. They'll show up at the ATM.