The next time you withdraw cash, the machine in front of you may be talking to software written in a language invented in 1959.
That language is COBOL — short for Common Business-Oriented Language. It was built so businesses could write programs that read more like English than like math. Decades later, a lot of the world’s money still moves through it.
The Big Idea: COBOL still supports 95% of ATM transactions worldwide — and OpenAI’s own test says AI can now rewrite old COBOL with near-98% accuracy. Experts still have to check the work.
How Much Still Runs on COBOL?
According to figures cited by OpenAI in an October 2026 post, COBOL still supports:
- 95% of ATM transactions worldwide
- 80% of in-person credit card transactions
- 40%+ of online banking systems
Those numbers come from OpenAI’s write-up (which points to industry reporting), not from an independent audit OpenAI ran itself. Treat them as widely cited estimates, not a new census of every bank.
Beyond banking, COBOL has long shown up in other large, record-heavy systems — including government benefit programs such as Social Security and parts of airline reservation software. Those uses are common industry knowledge; OpenAI’s post does not specifically name Social Security or airlines.
Why Is Updating It So Hard?
Much of this software is decades old. The original authors may be gone. Documentation can be incomplete. And the “business rules” — how interest is rounded, how a missing record is handled — often live only inside the code.
One small mistake in a rewrite can change payments, balances, or important records. That is why banks and agencies move carefully.
Modernizing COBOL is less like painting a house and more like rebuilding the foundation while people still live inside.
What Is OpenAI Testing?
On October 19, 2026, OpenAI published work on low-resource programming languages — languages with less public code and documentation for AI to learn from than Python or JavaScript. COBOL is one of them.
OpenAI built a COBOL migration evaluation: AI models try to rewrite old COBOL programs into modern languages (C++, Java, Kotlin, Python, or Rust) without changing what the programs do.
That last part is the hard part. COBOL’s fixed-point math, for example, often drops extra decimal digits unless rounding is asked for. A rewrite that rounds instead can look fine and still produce different money results.
OpenAI’s test only counts a migration as successful if hidden tests show the new code’s outputs match byte for byte.
How Well Did the AI Do?
On OpenAI’s own test (not an outside lab’s):
| Test | Result |
|---|---|
| GPT-6.1 Sol — COBOL migration, low reasoning effort | 86.6% accuracy |
| GPT-6.1 Sol — COBOL migration, higher reasoning effort | 97.9% accuracy — almost 98% |
| GPT-6 Astra — MultiLangBench coding test, 20 languages (context) | 87.4% average |
OpenAI also reported big gains on a separate coding test, MultiLangBench, where GPT-6 Astra averaged 87.4% across 20 languages. Useful context, but the ATM story turns on the COBOL migration numbers above.
Important caveat: these are vendor-reported scores on OpenAI’s own evaluation. And OpenAI itself says applying this to production software still requires expert review and testing against the system’s real business rules and dependencies. Near-98% on a benchmark is not “hand the bank to the AI and walk away.”
Frequently Asked Questions
What is COBOL?
COBOL — Common Business-Oriented Language — is a programming language from 1959 built so businesses could write programs that read more like English than like math. It still powers much of the world’s banking infrastructure.
How much still runs on COBOL?
According to figures cited by OpenAI: 95% of ATM transactions worldwide, 80% of in-person credit card transactions, and more than 40% of online banking systems. Treat these as widely cited estimates, not an independent audit.
Can AI really rewrite COBOL?
On OpenAI’s own migration test, GPT-6.1 Sol reached 97.9% accuracy at higher reasoning effort — but only when hidden tests confirmed the rewritten code’s outputs matched byte for byte. OpenAI itself says production use still requires expert review against the system’s real business rules.
Why This Matters
AI isn’t just inventing new technology. It could help upgrade the old systems we already depend on every day — as long as humans still check the work.
If you bank, fly, or collect a benefit, some of that reliability still sits on code from before most of us were born. The practical lesson from OpenAI’s test is simple: AI can accelerate the rewrite. Experts still have to sign off before anyone presses “go” on the systems that move real money.
Disclaimer: This article and accompanying media are for informational and educational purposes only and do not constitute legal, financial, tax, investment, or technical advice. While content is compiled from public sources, analyst reports, and company statements, pricing, specifications, benchmarks, and availability fluctuate over time and by region. All information is provided “as is” without warranty of any kind. Always conduct independent due diligence before purchasing software, making investments, or modifying business workflows.
Affiliate Disclosure: Some links in this article are partner or affiliate links. If you make a purchase, create an account, or start a subscription through them, we may receive a commission at no additional cost to you. All third-party tools and services are used strictly at your own discretion and risk.

