GPT-6 reaches 97.9% accuracy on COBOL migration tasks

The headline accuracy figures obscure a more important structural detail: OpenAI introduced a purpose-built variant, GPT-6.1 Sol, rather than relying on the general-purpose model. That specialization choice signals this is a deliberate product wedge into enterprise contracts, not a capability spillover from general training.

This story connects directly to OpenAI's October 19th research post on low-resource programming languages, which introduced the MultiLangBench framework and flagged COBOL migration as a target workflow. That piece was the research announcement; this is the capability validation. Together they suggest a coordinated release strategy: publish the evaluation framework, then demonstrate the model clearing it. Worth noting that GPT-5.6 Terra appears here as a comparison point, while GPT-5.6 Sol appeared in the Taste-Bench coverage from October 18th scoring only 60 percent on long-horizon decision tasks. The variant naming is getting complex enough that buyers will need clarity on which model applies to which workflow before committing to enterprise deployments.

Watch whether a major financial institution or government agency announces a COBOL migration pilot using GPT-6.1 Sol within the next two quarters. A named production contract would confirm this is a real procurement story; continued silence would suggest the benchmark gains haven't yet translated to procurement conversations.