OpenAI Says AI Is Getting Better at Fixing Old Bank Software

OpenAI says its AI models are now better at building and moving software written in COBOL, Fortran and Visual Basic. Engadget
The claim appeared in a blog post titled 'Advancing coding in low-resource programming languages' published on October 19, 2026. OpenAI OpenAI calls COBOL a low-resource programming language, meaning there are few public examples for AI to learn from.
OpenAI states that 40% of the world's banks still rely on COBOL systems. It aims to help banks update older software their operations still depend on.
To test progress, OpenAI built a test called MultiLangBench. It asks the model to build a small database that works with SQLite, a common system for storing data, in a set older language.
On that test, GPT-6 Astra passed 87.4% of checks on average across 20 languages, compared to 26.8% for GPT-5.6 Sol. The gap is large. COBOL remains the hardest language in MultiLangBench, with GPT-6 Astra passing 72% of COBOL checks.
OpenAI also tested moving COBOL programs to C++, Java, Kotlin, Python or Rust without changing what the programs do, like translating an instruction manual while keeping the meaning the same. GPT-6.1 Sol achieved 97.9% accuracy on COBOL migration with high reasoning effort and 86.6% accuracy with low reasoning effort. Reasoning effort means how much extra time and computing the model uses.
OpenAI said applying its behavior-preserving migration capabilities to production software still requires expert review and testing against business rules and dependencies.
Looking at what this means for engineering teams, that reasoning-effort split matters for daily work. High effort buys more than eleven points of accuracy in OpenAI's numbers. That gain maps directly to inference cost, latency and review load, meaning computing cost, waiting time and human checking. Teams will have to decide where to spend deliberation. Bulk triage of a large estate may tolerate lower effort with heavier human sampling. Cutover candidates will not.
The broader context here is that migration accuracy and production readiness are different thresholds. A model can preserve observable behavior on a program in isolation and still fail on the system around it. Business rules live outside the source file. So do job control, transaction monitors, batch schedules, data encodings and undocumented dependencies accumulated over years. OpenAI's caution about expert review and testing points to that boundary. The work shifts from rewriting code to proving the new version works under real operating rules.
In my view, there is reason for measured optimism. Low-resource languages have been a weak point for code assistants because public examples and modern tooling are thin. Gains there expand what automation can usefully touch inside banks and other institutions tied to legacy estates. The path is unlikely to be wholesale replacement. It looks more like incremental extraction, with models drafting the target-language equivalent and engineers owning validation. If that pattern holds, the value is less in removing COBOL at speed and more in making modernization continuous without freezing the systems that still run the business.


