OpenAI has reported progress in using its AI models to migrate legacy software written in COBOL to modern languages like C++, Java, Python, Kotlin, and Rust. In internal tests, the GPT-6.1 Sol model achieved 98 percent behavior preservation accuracy at high reasoning effort, compared to 87 percent at low effort.
The company stated that successful migration requires all tests to pass, as the goal is to modernize systems without altering original software functionality. OpenAI also noted improvements in its models' ability to build database software using various programming languages, including COBOL.
Anthropic is also targeting legacy systems, having released a code modernization playbook in February. The company promotes its coding agent, Claude Code, as a tool for migrating COBOL to modern languages. Following this announcement, IBM shares experienced a decline.
Experts emphasize that AI-driven code rewriting is not equivalent to replacing entire systems. Challenges include the massive scale of legacy codebases, complex technical integrations, and a lack of documentation. Academic researchers noted that current AI benchmarks remain far from real-world modernization tasks.
The reliance on COBOL remains significant, with IBM reporting that it supports 40 percent of online banking systems, 80 percent of in-person credit card transactions, and 95 percent of ATM transactions. Estimates suggest COBOL inefficiencies cost the US GDP around $105 billion annually.
