# OpenAI Enhances AI Models for Legacy COBOL Banking Migrations New benchmarks show improved accuracy in translating aging banking code to modern languages. - **Source:** [Engadget](https://www.engadget.com) - **Published:** 2026-10-20 - **Topic:** [Data & AI](https://brifing.si/c/data-ai) - **Original article:** - **On Brifing:** https://brifing.si/s/26cd5d05-f572-4d7d-b88b-b38858cfbe78 ## Legacy System Migration Challenges Approximately 40 percent of global banks still rely on COBOL systems for core operations. Migrating these legacy architectures to modern languages typically costs millions of dollars and requires significant time. As the workforce proficient in COBOL ages out, financial institutions face an urgent need to modernize their infrastructure within the coming years. ## Enhanced Model Capabilities OpenAI has optimized its models to better handle older programming languages including COBOL, Fortran, and Visual Basic. This development is directly informed by feedback from enterprise customers and developers who identified specific struggles in legacy code processing. The goal is to facilitate smoother transitions while maintaining the integrity of critical banking software. ## MultiLangBench Evaluation Suite OpenAI introduced MultiLangBench to evaluate model performance across 20 different programming languages. The suite tests the ability to create SQLite-compatible database engines. Results show GPT-6 Astra achieved an 87.4 percent pass rate, significantly outperforming the 26.8 percent pass rate of GPT-5.6 Sol. COBOL remains the most challenging language, with a 72 percent success rate. ## Behavior-Preserving Code Translation Successful migration requires preserving specific program behaviors, such as rounding numbers with excess decimal digits or handling missing records. OpenAI tested its models by migrating COBOL programs to C++, Java, Kotlin, Python, or Rust. GPT-6.1 Sol reached 97.9 percent accuracy in these tasks when utilizing high reasoning effort during the translation process. ## Reasoning Effort Impact The accuracy of code translation is heavily dependent on the reasoning effort applied by the model. While GPT-6.1 Sol achieved 97.9 percent accuracy with high effort, this figure dropped to 86.6 percent under low reasoning effort. This indicates that complex banking logic requires substantial computational resources to ensure the output remains functionally equivalent. ## Strict Success Criteria OpenAI emphasizes that a migration is only considered successful if all tests pass. Because banking tasks involve high-stakes financial data, the margin for error is extremely narrow. The models must demonstrate perfect adherence to existing business rules and dependencies to be considered viable for real-world application in production environments for financial institutions. ## Expert Review Requirement Despite improvements in model capabilities, OpenAI maintains that production software migration still requires expert human review. Automated tools serve as a bridge for maintaining and modernizing older systems, but they cannot replace the necessity of testing against established business rules. Human oversight remains a critical component of the final deployment phase. ## Strategic Industry Implications The ability to automate legacy code migration addresses a major bottleneck in the financial sector. By reducing the reliance on a dwindling pool of COBOL experts, banks can accelerate their digital transformation efforts. This shift represents a broader trend of leveraging generative AI to manage technical debt within large-scale, mission-critical enterprise systems. ## Sources covering this story - https://www.engadget.com/2291457/openai-says-its-models-can-translate-cobol/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._