OpenAI Says GPT-6.1 Sol Preserves 98% of COBOL Behavior in Migration Tests
Updated
Updated · The Verge · Oct 26
OpenAI Says GPT-6.1 Sol Preserves 98% of COBOL Behavior in Migration Tests
3 articles · Updated · The Verge · Oct 26
Summary
OpenAI said its latest models made substantial gains in translating legacy code such as COBOL into modern languages, with GPT-6.1 Sol reaching 98% behavior preservation at higher reasoning effort.
The company framed that as progress for organizations trying to modernize critical systems without changing what the software does; at low reasoning effort, GPT-6.1 Sol scored 87%, versus 21% for GPT-6 Sol and 5% for GPT-5.6 Sol.
COBOL remains deeply embedded in infrastructure—IBM says it underpins more than 40% of online banking, 80% of in-person card transactions and 95% of ATM transactions—making accurate migration a potentially huge market.
Experts said the benchmarks still fall far short of real-world modernization, where systems can span hundreds of millions of lines, multiple languages, data formats and external integrations.
Researchers said AI may speed parts of the work, but human verification and stronger testing remain essential because even minor behavior changes can be unacceptable in critical systems.
Could replacing ancient COBOL systems with AI-generated code introduce more catastrophic vulnerabilities into critical infrastructure than the legacy software itself?
If AI seamlessly translates 60-year-old banking code, who takes the blame when a hidden logic error drains millions of accounts overnight?