Banks Face $100 Billion COBOL Crisis as Programmers Retire — Can AI Fill the Gap?
OpenAI benchmarks show 97.9% accuracy translating COBOL — but regulators still require human sign-off
The global banking system runs on code written before the moon landing. Forty percent of banks still process transactions on COBOL systems, and the programmers who maintain them are vanishing faster than institutions can replace them. Migration costs run from millions to hundreds of millions per bank — a collective bill that could exceed $100 billion industry-wide.
OpenAI says its latest models can now help. The company released benchmarks showing GPT-6 Astra passing 87.4 percent of tests across 20 programming languages in a new evaluation suite called MultiLangBench. Its predecessor, GPT-5.6 Sol, managed just 26.8 percent. But COBOL remains stubborn: Astra scored only 72 percent.
Migration is where the numbers look better. When asked to translate COBOL into modern languages — C++, Java, Kotlin, Python, or Rust — while preserving exact behavior, GPT-6.1 Sol hit 97.9 percent accuracy at high reasoning effort. At low effort, accuracy dropped to 86.6 percent.
Preserving behavior means more than syntax. COBOL rounds numbers differently than modern languages. It handles missing records in ways that, if mistranslated, could misstate account balances by pennies — multiplied across billions of transactions. OpenAI acknowledges that “all tests pass” is the only acceptable standard for production banking software.
The demographic cliff is real. The average COBOL developer is over 60. Universities stopped teaching it decades ago. Banks have been “five years from migration” for twenty years. AI-assisted translation could finally break the logjam — but only if regulators accept machine-translated code without line-by-line human verification.
OpenAI developed MultiLangBench after feedback from enterprise customers and developers. The benchmark tasks models with building SQLite-compatible database engines in each target language. Results vary wildly by language family. COBOL’s columnar structure, fixed-format lines, and decimal arithmetic have no clean analogs in C-style languages.
Expert review remains mandatory. “Applying it to production software still requires expert review and testing against the system’s business rules and dependencies,” OpenAI wrote. No regulator has approved AI-translated core banking code for production use.
