OpenAI says GPT-6.1 Sol now migrates COBOL correctly 86.6% of the time, up from 5.4%
OpenAI says GPT-6.1 Sol can now move old COBOL programs to modern languages with far fewer mistakes. In two new tests OpenAI built and ran itself, GPT-6.1 Sol passed 86.6% of COBOL migrations at low reasoning effort. GPT-5.6 Sol managed 5.4%, and the new model cost 39% less to run.
OpenAI's title card for GPT-6.1 Sol, the model behind the COBOL results. Image: OpenAI
At higher effort, the score reaches 97.9%. OpenAI says that is close to the most the test can measure.
COBOL still runs banks, and rewriting it can quietly change the numbers
COBOL, Fortran and Visual Basic have far less public code for models to learn from than Python or JavaScript. Yet COBOL still handles much of the world's banking, and Fortran runs weather forecasts.
OpenAI's test asks a model to rewrite COBOL programs in C++, Java, Kotlin, Python or Rust. Hidden tests then check that every output matches the original byte for byte, and a migration only counts if all of them pass.
OpenAI gives one example of the trap. COBOL drops extra decimal digits unless told to round, so a rewrite that rounds can run without errors and still produce different financial results.
OpenAI's title card for GPT-6 Sol and Luna. Luna is the cheaper model in the SQLite test. Image: OpenAI
GPT-6 Astra passed 87.4% of tests when building a database in 20 languages
The second test, MultiLangBench, asks a model to build a database engine that behaves like SQLite from scratch, in 20 programming languages. OpenAI checks each result against the real SQLite, using tests adapted from the SQLite test suite.
GPT-6 Astra averaged 87.4% across the 20 languages, up from 26.8% for GPT-5.6 Sol. It scored above 83% in 19 of them. COBOL stayed the hardest at 72.1%, up from 10.9%. The cheaper GPT-6 Luna reached 40.3% at an estimated 78% lower API cost than GPT-5.6 Sol. OpenAI has put the test on GitHub so developers can see what it checks.
Artwork from OpenAI's Codex page. OpenAI invites developers to try the new models on their own code. Image: OpenAI
OpenAI built both tests and compared only its models with each other
Both benchmarks are new and built by OpenAI, and no rival models appear in the results. OpenAI also says any real migration still needs expert review and testing against the business rules it has to preserve.
The results follow OpenAI's earlier Ironclad evaluation of GPT-6 Astra, which also came with OpenAI's figures alone. For teams sitting on decades of COBOL, the 97.9% figure is a reason to run their own tests, not to skip them.
