OpenAI says GPT-6.1 Sol delivers substantial improvements in legacy code migration, including converting COBOL programs into newer programming languages. In the company’s evaluation, the model reached 97.9% accuracy at higher reasoning effort.
The findings accompany broader improvements in programming languages with relatively limited training material. However, these evaluation results do not guarantee reliable performance across every production system.
Older Languages Still Support Essential Systems
Languages such as COBOL, Fortran and Visual Basic continue to support important software, despite receiving less attention than newer alternatives.
According to OpenAI, COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions. Meanwhile, Fortran remains relevant to weather forecasting and nuclear reactor safety analysis.
These languages present a particular challenge for AI development. Compared with Python and JavaScript, they offer less publicly available code and documentation for models to learn from.
Nevertheless, organisations still need to maintain these systems and, in some cases, move their functionality to newer platforms.
GPT-6.1 Sol Improves COBOL Conversion Results
OpenAI tested whether its models could translate COBOL programs into C++, Java, Kotlin, Python or Rust while preserving their behaviour.
At low reasoning effort, GPT-6.1 Sol achieved 86.6% accuracy. By comparison, GPT-5.6 Sol scored 5.4% under the same reasoning setting.
OpenAI also reported that GPT-6.1 Sol completed the evaluation at 61% of the older model’s API cost. That represents a 39% reduction for the measured workload.
At higher reasoning effort, GPT-6.1 Sol reached 97.9% accuracy. However, the cost comparison supplied for the low-effort setting does not establish the cost of that higher-effort result.
These findings suggest stronger performance on the tested legacy code migration tasks, rather than proof that companies can automate entire modernisation projects without oversight.
Separate Benchmark Tests Code Generation Across 20 Languages
OpenAI also introduced MultiLangBench to assess programming performance across 20 languages. The benchmark asks models to build a SQLite-compatible database engine from scratch.
This evaluation measures code generation, making it distinct from the COBOL migration test.
On MultiLangBench, GPT-6 Astra achieved an average test pass rate of 87.4%, compared with 26.8% for GPT-5.6 Sol.
The company reported improvements in every language it tested. For example, COBOL results rose from 10.9% to 72.1%, while Fortran increased from 21.4% to 83.5%.
Similarly, OCaml improved from 25.6% to 85.8%. Visual Basic rose from 27.5% to 89.4%.
Production Migrations Still Need Expert Validation
The reported gains could help organisations update older software and reduce some of the work involved in translating code.
However, preserving behaviour remains central to a successful migration. A program must continue producing the expected results after its implementation changes.
Therefore, production projects still require expert review and thorough testing. OpenAI’s evaluations show progress on defined tasks, while real deployments must establish that the converted software meets their specific requirements.
0 responses to “OpenAI Reports Major Gains in Legacy Code Migration With GPT-6.1 Sol”