OpenAI breaches legacy banking systems with an accuracy leap in COBOL code migration

Banks and government agencies worldwide retain thousands of systems that were written decades ago in poorly documented programming languages, and they have long been difficult to update because of the risk of disrupting the accumulated business logic they contain. On 19 October 2026, OpenAI announced a technical initiative that directly targets this dilemma by enhancing its models' capabilities to build and migrate software in resource-constrained and data-scarce languages, foremost among them COBOL, Fortran, Visual Basic and OCaml.

When we examined the figures announced by the company, we found a striking shift in model performance on legacy code. In COBOL migration tests, the GPT-6.1 Sol model achieved an accuracy of 86.6 percent under low-inference settings, far outpacing the 5.4 percent recorded by the GPT-5.6 Sol model, while reducing cost by up to 39 percent. Raising the inference settings to their highest levels brought the same model’s migration accuracy to 97.9 percent, narrowing a programming gap that has historically made moving these systems one of the most complex challenges for development teams.

The assessment did not stop at a single language; the company released the MultiLangBench framework and its full test suite on GitHub to cover twenty programming languages with detailed results. In this benchmark, the GPT-6 Astra model recorded a pass rate of 87.4 percent across the twenty languages, up from 26.8 percent for the GPT-5.6 Sol model. In contrast, the GPT-6 Luna model showed an explicit economic trend, achieving a pass rate of 40.3 percent while costing about 78 percent less than earlier models, giving institutions room to balance high accuracy against spending.

This jump in success rates does not mean unleashing automation in sensitive work environments without oversight.OpenAI clearly warned that adopting these outputs in actual production software remains conditional on specialized human verification and rigorous testing that ensures the code conforms to complex business rules and existing operational dependencies. The fundamental difficulty in languages such as COBOL does not stem solely from the scarcity of training models compared with contemporary languages like Python; it also arises from layers of banking and regulatory decisions accumulated over successive generations, so that any minor error in translating transaction logic could trigger widespread financial disruption.

For the banking and governmental sectors in the Gulf and Egypt, this development carries a highly significant operational dimension. Major financial institutions in the region rely on core banking systems whose original code was written decades ago, and the migration concern has remained postponed due to high costs of specialized consulting and the scarcity of programmers capable of deciphering legacy languages. If you lead the technology function in a regional bank, your practical step does not begin with an immediate migration but with understanding the shift in computing costs; with costs dropping by as much as 78 percent and a maturing grasp of obsolete languages, financial modernization projects become feasible without draining budgets. However, this shift moves the skill focus: the requirement is no longer merely a developer who can write script commands, but an enterprise-architecture engineer who can audit business logic and rigorously verify operational integrity before approval.

True success is measured by how well an organization can conduct rigorous verification before adopting new code.Providing the MultiLangBench framework gives engineering teams in the region an independent tool to test models on local codebases, pinpoint error margins precisely before touching live data, turning artificial intelligence from a superficial interface into a deep instrument for addressing accumulated technical debt.