I was on a late-night support call when a teller confessed the branch’s withdrawal logic ran on code written before the moon landing. You could hear the stress in their voice as they described a decimal error that only appears at month-end. That moment made me stop and ask: what if AI could speak the language of these old machines?
An ATM in Midtown still runs COBOL code installed when Eisenhower was in the Oval Office. That fact explains why the financial system feels fragile even as fintech apps thrive.
I follow these seams because you should know what keeps your money moving. COBOL—common business-oriented language—still handles roughly 95% of ATM transactions. Fortran, born the same decade, still underpins reactor simulations. The people who learned those languages are retiring, and their tacit knowledge is leaving with them.
Meanwhile, younger engineers turn to assistants such as OpenAI’s Codex and Anthropic’s Claude Code. Those tools were trained mostly on modern languages, widening a gap between new workflows and the legacy stacks running banks, utilities, and government systems.
Legacy code is a fossilized engine that still powers whole neighborhoods. If it breaks, fixes are expensive, slow, and politically painful.
At a developer lab, OpenAI fed its models a 1970s COBOL payroll program and logged every output. The experiments were aimed at whether machine learning can safely move behavior, not just syntax.
OpenAI announced that GPT-6 Astra and GPT-5.6 Sol show “substantial improvements” in migrating old languages into modern ones. They built two benchmarks to measure progress. MultiLangBench tests whether a model can recreate an entire program—SQLite—from scratch in a given language. GPT-6 scored about an 87% average pass rate on that challenge.
The more delicate test was migrating COBOL programs into languages such as C++, Java, Kotlin, Python, or Rust without changing behavior. Small differences in rounding rules, date handling, or fixed-point math can ruin payroll or settlements. OpenAI reports GPT-6.1 Sol, when run with higher reasoning effort, scored very close to 100% on that migration test.
These aren’t perfection claims made in a vacuum. The work was “guided by feedback from developers and enterprise customers,” which is shorthand for real cases and edge rules—those quirks that make old systems hard to replace.
Can AI modernize COBOL systems?
Yes—with caveats. Models can translate syntax and preserve many semantic behaviors rapidly, turning years of manual effort into months or quarters. But that speed depends on careful vetting: test suites, regression checks, and human review remain mandatory.
How does OpenAI migrate legacy code without changing behavior?
They combine targeted benchmarks, large-model reasoning, and developer feedback loops. The COBOL tests measure not just compile-and-run but equivalence in financial behavior—rounding, overflow handling, and business rules—then score pass rates across languages.
Is it safe to trust AI for banking or nuclear simulations?
Trust is a process, not a product. Models can automate the mechanical parts of migration, but institutions must own verification, logging, and rollback procedures. No single model should be allowed to deploy unsupervised into production systems that affect lives.
At a trading desk, the market reacted the moment AI companies mentioned COBOL modernization. That reaction revealed how sensitive legacy modernization is to perception.
When Anthropic said Claude Code could speed COBOL modernization from years into quarters, IBM—a major mainframe vendor—saw a 13% share drop reported by Bloomberg the same day. That market move signals two things: incumbents fear disruption, and investors believe AI can lower the cost of a historically expensive process.
I want to be clear with you: faster migrations mean fewer payroll outages and fewer single-person failure points. They also mean automated tools will be touching systems that were never designed for automated change. Which brings me to a practical point.
AI acts as a locksmith opening back doors without blowing the safe. You need strict change controls, simulation environments, and an audit trail that proves behavior parity before you flip a production switch.
What you should watch for: test coverage that mirrors financial behavior, teams that pair legacy experts with AI-guided engineers, and clear governance over automated rewrites. OpenAI, Anthropic, Claude, Codex, and vendors such as IBM are reshaping how this work gets done—but the technical and political work of modernization still lives with humans.
If models can rewrite the nervous system of a bank, who decides when the scalpel is sharp enough to operate?