OpenAI's New AI Can Finally Update Ancient Bank Code

OpenAI announced that its newest AI models can now translate COBOL—a 65-year-old programming language that still runs most banking and government systems—into modern code. In tests, the AI preserved the original behavior 98% of the time when given more time to think. That's a big jump from older models that scored as low as 5%. This matters because COBOL is everywhere: it handles 80% of credit card transactions and 95% of ATM withdrawals, and the IRS and Social Security rely on it. The cost of maintaining these outdated systems is estimated at $105 billion a year in the US alone.

But don't expect your bank to switch overnight. Experts say rewriting code is not the same as replacing a whole system. Critical systems are too important to risk breaking, and many of the programmers who understand them have retired. OpenAI admits legacy code is 'still far from a solved problem.' The company is working on better testing to ensure migrations are safe.

Anthropic, another AI company, is also targeting this market with its Claude Code tool. When Anthropic announced its COBOL modernization playbook in February, IBM's stock dropped, showing investors see AI as a threat to traditional IT services. However, Anthropic didn't provide performance numbers, so it's unclear how well its tool works.

The takeaway: AI is getting better at understanding old code, which could eventually save money and reduce the risk of system failures. But for now, it's a promising step, not a cure. Your bank's COBOL systems aren't going away anytime soon—but AI might help them run more smoothly in the future.