GPT-6.1 Sol Scores 97.9% On OpenAI's COBOL Test - Dataconomy
OpenAI said its newest models have become much better at writing and converting code in programming languages such as COBOL. GPT-6.1 Sol scored 86.6% on a test that measures how well it can convert COBOL programs to other languages.
GPT-6.1 Sol achieved this score using a low level of reasoning effort. By comparison, GPT-5.6 Sol scored 5.4%. GPT-6.1 Sol also cost 61% as much to use through the API. When given more reasoning effort, GPT-6.1 Sol reached 97.9%. OpenAI said this result brought the test close to its maximum possible performance.
OpenAI created and ran both tests announced on October 19, 2026. No independent group has checked the results.
The company said that using these abilities in real-world software still requires experts to review the code and test it against the software's business rules and connections to other systems.
The COBOL conversion test asks models to change programs into C++, Java, Kotlin, Python or Rust without changing what the programs do. Hidden tests check whether the new programs produce results that match the original programs exactly, down to every byte. A conversion counts as successful only when it passes every test.
OpenAI said COBOL's way of handling decimal numbers makes this task especially difficult. COBOL uses fixed-point arithmetic, which means it removes extra decimal digits unless the program is told to round the numbers. If a rewritten program rounds numbers when the original did not, it may run without errors but still produce different financial results.
The second test, MultiLangBench, asks models to build a database engine from scratch that works with SQLite, using each of 20 programming languages. The results are checked against the original SQLite software with tests adapted from SQLite's own test suite.
When using the highest level of reasoning effort, GPT-6 Astra passed an average of 87.4% of the MultiLangBench tests, compared with 26.8% for GPT-5.6 Sol. Astra scored above 83% in 19 languages. COBOL was the hardest language, but its score still rose from 10.9% to 72.1%.
The less expensive GPT-6 Luna passed an average of 40.3% of the tests. Its estimated API cost was $0.77 per sample, compared with $3.48 for GPT-5.6 Sol. That made GPT-6 Luna about 78% cheaper.
OpenAI describes COBOL, Fortran and Visual Basic as programming languages with fewer available resources because there is less code and documentation to train models on than there is for languages such as Python and JavaScript. Developers and business customers told OpenAI that coding models often have trouble working with these languages.
The company cited figures showing that COBOL supports more than 40% of online banking systems and 95% of ATM transactions. It also said Fortran is used for weather forecasting and for checking the safety of nuclear reactors.
COBOL was created in 1959. According to ITHardware, it rose from 23rd place in October 2025 to 16th place in the TIOBE programming language popularity index. Companies still rely on COBOL systems, even as the number of engineers who can maintain them is shrinking.
OpenAI is sharing the MultiLangBench setup, test suite and results for each programming language on GitHub.