OpenAI's GPT-6 takes on ancient banking code

Every time you swipe a card or tap cash from an ATM, there is a very high chance the transaction is quietly routed through code written before your parents met.

Over 40% of online banking systems, 80% of in-person credit card swipes, and 95% of ATM transactions still rely on COBOL. Meanwhile, systems like weather forecasting and nuclear safety analysis lean on Fortran. The engineers who built them are mostly retired, documentation is often non-existent, and AI coding assistants have historically choked on these languages due to a lack of training data.

Now, OpenAI says its newest crop of frontier models has finally figured out how to write and modernize legacy code.

Building database engines in 20 languages

To prove its models actually understand rare languages rather than just spitting out boilerplate, OpenAI designed a new benchmark called MultiLangBench.

The task is brutal: write a functional SQLite-compatible database engine from scratch across 20 different languages, validated against SQLite's official test suite to ensure the resulting database files are identical.

GPT-6 Astra blew past previous standards, hitting an average pass rate of 87.4% across the 20 languages, compared to just 26.8% for GPT-5.6 Sol. Astra cleared over 83% in 19 out of 20 languages, with COBOL proving the toughest nut to crack at 72.1%.

Other historical languages saw massive jumps too. Fortran vaulted from 21.4% to 83.5%, Visual Basic jumped from 27.5% to 89.4%, and OCaml surged from 25.6% to 85.8%.

For cost-sensitive teams, GPT-6 Luna delivered a 40.3% average pass rate while slashing API costs by roughly 78% down to $0.77 per sample, outperforming the previous flagship at a fraction of the price.

COBOL migration jumps from 5% to 98%

Writing new legacy code is only half the battle. What banks and enterprises desperately need is safely converting aging COBOL programs into modern languages like Rust, Python, Java, Kotlin, or C++.

The real challenge lies in obscure language quirks. For instance, COBOL's fixed-point arithmetic silently truncates excess digits unless rounding is specifically declared. An AI that accidentally rounds those numbers might generate clean code that runs without errors, while silently botching a bank balance.

To test this, OpenAI built a dedicated migration benchmark requiring byte-for-byte output parity against hidden tests.

GPT-6.1 Sol delivered a near-perfect run. At low reasoning effort, it jumped to an 86.6% migration accuracy at 39% lower cost, compared to a miserable 5.4% from GPT-5.6 Sol. Cranking the reasoning effort higher pushed accuracy to 97.9%, virtually maxing out the test.

Why it matters

This is a massive breakthrough for the least glamorous, highest-stakes corner of enterprise software.

Trillions of dollars in global commerce flow through systems that no active employee fully understands. Rewriting these architectures from scratch has historically been a corporate career suicide move due to the catastrophic risk of subtle logic bugs. Pushing code migration accuracy from 5.4% to nearly 98% changes the risk calculus entirely.

We think OpenAI's decision to target low-resource and legacy languages is extraordinarily smart. While rivals fight over incremental gains in Python and JavaScript, tackling COBOL and Fortran creates direct enterprise leverage where budgets are deep and alternatives are virtually zero.

That said, production reality will still demand extreme caution. An AI matching byte-for-byte behavior on isolated routines is impressive, but real-world mainframes are wrapped in decades of undocumented side effects and archaic hardware quirks. It gives modernization teams an incredible engine to start with, but nobody should be pushing these rewrites straight to main without rigorous human auditing.