OpenAI has improved the latest models for low-resource programming languages, including COBOL, Fortran and Visual Basic, The Decoder reports. According to the paraphrase from the publication, on the MultiLangBench test GPT-6 Astra showed a result 87,4% for 20 languages against about 38% for GPT-5.6 Terra.
For porting programs to COBOL in the material, the accuracy is up to 97,9% at GPT-6.1 Sol. This may be important for updating critical financial infrastructure where old languages continue to be used.
The data are based on metadata and the paraphrase of a single independent source; primary OpenAI material is also presented only as a brief description. The package does not contain details about testing methodology, task composition, or independent verification of results.
editorial commentary
Why it matters
Practical implication likely — greater attention to application of models during maintenance and porting of old software systems. The next checkable signal will be the publication of a full assessment methodology and independent verification of results. Substantial uncertainty is linked to the fact that currently only metadata and a paraphrase of one independent source are available.