Artificial intelligence achieves 97.9% accuracy in COBOL code migration, OpenAI says

OpenAI announced on Monday, 19, that its GPT-6.1 Sol model successfully rewrote COBOL programs with a 97.9% success rate in internal tests. The experiment tested the AI’s ability to convert legacy code into other languages without changing the original logic, while preserving financial and administrative functions.

Why old code is still in use

Legacy systems are programs inherited from previous generations of technology that remain in use because they contain business rules accumulated over decades. Replacing them requires not only technical expertise, but also an understanding of operational history and integration with other systems.

As early as 2014, there was discussion about converting COBOL applications to Java to simplify maintenance and enable the use of modern computing environments, while preserving the original logic. The shortage of qualified professionals exacerbates the problem. A 2025 report by the GAO, the U.S. government’s oversight agency, highlighted difficulties in finding experts capable of maintaining old programs, citing as critical examples Treasury systems that rely on COBOL and Assembly.

Results of OpenAI’s tests

In addition to COBOL migration, OpenAI introduced MultiLangBench, a public evaluation available on GitHub. In this evaluation, the AI receives a database and an instruction, executes the task, and compares the result with the behavior of the original SQLite, verifying whether the program works correctly, not just whether the code text looks appropriate.

The tests released were created by the company itself and do not represent independent validation. OpenAI warns that small differences can be significant: a conversion may round values that the original program simply truncated, resulting in different calculations even when the code appears to work.

Competitors in the modernization market

Anthropic also offers Claude Code, a tool designed to analyze old programs, document their rules, and assist in converting them to languages such as Java and Python. IBM, a traditional provider of these environments, offers AI capabilities to explain and reorganize code, noting that modernization involves integration, data, and security. For some organizations, maintaining COBOL may be feasible if combined with new software layers.

Workforce and cost challenges

In Brazil, a study conducted by TQI, Fundação Dom Cabral, and Integratum found that 76% of the 51 organizations surveyed allocate at least 70% of their technology budget to maintaining existing systems. The sample focused on large companies, where the work of preserving legacy code is more intensive.

Dependence on old code is also evident in critical sectors. For example, U.S. Treasury programs still operate using COBOL and Assembly, making replacement difficult without qualified specialists.

Implications for the industry

The ability of AI to reinterpret legacy code can reduce costs and accelerate the transition to modern platforms, but still requires human verification to ensure that results remain consistent with original requirements. Companies such as IBM and Anthropic already integrate testing into their modernization processes, emphasizing that automation complements, but does not replace, careful analysis of legacy systems.