OpenAI evaluated migrations of COBOL programs to C++, Java, Kotlin, Python and Rust, with the goal of preserving program behavior.
GPT-6.1 Sol reached 97.9% accuracy at high reasoning effort, compared with 86.6% at low effort.
The key constraint is validation: OpenAI counts a migration as successful only when every test passes.
The results show why model scores alone are not the same as a verified code migration.
As programmers of this language age and leave the job market, banks are forced to plan for a quick change to their systems. This improvement was made after feedback and observations from programmers and businesses.
How the new features were tested
To measure progress, OpenAI created an evaluation system called MultiLangBench, which tests the creation of database engines in various programming languages.
In tests, the GPT-6 Astra model achieved an average success rate of 87,4% across 20 languages. However, COBOL remained the most difficult language, with Astra achieving a rate of 72%.
READ ALSO
Yelp is putting a digital assistant to answer restaurant phone calls for you
Converting old programs into modern languages
The company also looked at how its systems can convert COBOL programs to C++, Java, Kotlin, Python or Rust, while maintaining the correct operation of the program.
In these tests, the GPT-6.1 Sol model achieved the highest accuracy, reaching 97,9% when operated with a high level of reasoning effort, while with a low level of effort it reached 86,6%.
OpenAI stressed that a conversion is only considered successful if it passes all the checks, which is essential for banking operations. At the same time, it was noted that using these capabilities in real programs always requires expert review and testing based on each business's rules and dependencies.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
