Can AI Solve the COBOL Code Crisis Holding Up Global Banking?

Behind every transfer, every ATM withdrawal, every credit card swipe, there is code written decades ago in COBOL — a programming language created in 1959 that still moves billions of dollars in operations every day. According to IBM, COBOL powers more than 40% of the world's online banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions. The language was designed for an era when computers occupied entire floors and programmers numbered in the hundreds. Today, running on mainframes that have become part of the economy's invisible infrastructure, COBOL is as ubiquitous as the air we breathe: you only realize it exists when it breaks.

And the problem is deepening. COBOL programmers are aging out of the workforce — the average age of a COBOL developer sits around 45 to 55 years old, and fewer and fewer young people are learning the language. The result is a silent crisis that has already begun: banks and financial institutions face a scenario where migrating legacy code to modern languages would cost millions — or hundreds of millions — of dollars, without guaranteeing the process wouldn't introduce catastrophic bugs into systems nobody dares touch.

OpenAI's announcement

On October 19, 2026, OpenAI announced significant advances in its AI models' ability to work with low-resource programming languages — COBOL, Fortran, and Visual Basic. The announcement came with the brutal results of two evaluations the company created to measure progress: MultiLangBench and the COBOL migration evaluation.

On MultiLangBench, which asks the model to build a SQLite-compatible database engine from scratch across 20 different programming languages, GPT-6 Astra achieved an average pass rate of 87.4% — a massive leap from the 26.8% of GPT-5.6 Sol. Across all 20 languages, results improved. For COBOL specifically, the model jumped from 10.9% to 72.1%. This means that, for the first time, an AI model can reliably write functional COBOL code.

But the real test — the one that matters to banks — was the migration evaluation. Here, the model receives an existing COBOL program and must migrate it to another language (C++, Java, Kotlin, Python, or Rust) while preserving exactly the same behavior. The crucial detail: the code must not just translate syntactically. It must maintain the business logic, including obscure details like how the program rounds numbers with excess decimal digits or handles missing records. For a banking system, a bug at this level could mean disappearing money, duplicate transactions, or incorrect regulatory reports.

GPT-6.1 Sol achieved 86.6% accuracy at low reasoning effort and 97.9% at high effort — a number so close to saturation that OpenAI's own article acknowledges the evaluation is "near saturation." The model also cost 39% less in APIs than the previous version. The numbers are impressive. The caution is equal.

The chasm between benchmark and production

In a controlled benchmark, 97.9% can sound close to perfect. In the real world of banking systems, 2.1% error means thousands of corrupted transactions. OpenAI itself acknowledges this in its article: "applying it to production software still requires expert review and testing against the system's business rules and dependencies."

The COBOL case is not just a technical problem — it's a problem of scale and systemic risk. When a bank decides to migrate, it is not about translating a to-do app. It is millions of lines of code spread across decades of mergers, acquisitions, and emergency updates. Each module depends on another. Each business rule was implemented differently across different systems. Antony Jenkins, former CEO of Barclays, told Sott.net that "it is immensely complex." For large financial institutions, "the problem of replacing old technology goes beyond a shrinking pool of experts."

The most apt analogy is thinking of COBOL migration as equivalent to rebuilding an airplane's engine mid-flight — and the airplane is a Boeing 747 carrying 300 passengers who don't know the engine is being swapped. Any migration failure could result in duplicated financial transactions, incorrect balances, false regulatory reports. And if the bug is discovered months later, after quarterly declarations have already been sent to regulators?

What comes next

What OpenAI demonstrated is not that COBOL migration is solved — it's that AI models are approaching rapidly a capability that, two years ago, seemed like science fiction. The leap from 5.4% to 86.6% in COBOL migration accuracy, in a single model generation, is faster than any analyst's prediction.

But there is a fundamental difference between a controlled benchmark and a banking mainframe. MultiLangBench tests isolated programs with known behavior. A bank has decades of legacy code where business rules are embedded in variable logic, comments in three different languages, and emergency patches applied 15 years ago that no one else understands. AI can document, translate, suggest migrations — but until the model can not just translate code but understand the business context that produced it, COBOL migration will remain a hybrid endeavor: AI doing the heavy lifting, humans doing the critical judgment.

The question is whether this hybrid approach will be fast enough to prevent skills crises in COBOL from causing systemic failures before migration technology matures. Or whether the moment the last veteran programmers retire will arrive before models are ready to touch the systems that sustain the global economy.

Sources: The Verge, OpenAI, Engadget

✓ Independent sources cross-checked and verified before publishing