OpenAI GPT-6 models post large gains on legacy programming benchmarks

By Splitfeed EditorialLast updated Oct 20, 2026, 17:10 UTC

On October 20, 2026, OpenAI published results showing its newest models performing substantially better on legacy and low-resource programming languages. GPT-6 Astra scored 87.4 percent on MultiLangBench, a new 20-language evaluation OpenAI introduced alongside the announcement, compared to roughly 38 percent for the previous generation model GPT-5.6 Terra. A separate, more focused test for COBOL migration tasks put GPT-6.1 Sol at 97.9 percent accuracy. The two figures come from two distinct evaluations, and the two models named are different releases within the GPT-6 family.

The practical framing OpenAI chose is financial infrastructure. Large banks and government agencies have run COBOL-based systems for decades, and the cost and risk of modernizing those systems have made migration a persistent problem in enterprise software. A model that can reliably translate or rewrite COBOL at near-perfect accuracy would reduce one of the main bottlenecks, though the gap between benchmark performance and production reliability is a recurring point of contention in the field. OpenAI designed and administered both evaluations itself, which is also worth noting, since self-reported benchmarks invite questions about methodology and selection effects that independent replication would help address.

None of the three camps have published specific reactions to this story yet. The Pro-AI camp would typically treat a jump from 38 percent to 87.4 percent in a single generation as strong evidence of continued capability scaling with direct economic value. The Anti-AI camp would likely raise concerns about whether proprietary benchmarks reflect real-world performance and whether deploying AI in critical financial systems introduces new categories of risk that outweigh the efficiency gains. The Middle Ground camp would probably welcome the progress on a genuine infrastructure problem while calling for third-party validation of the benchmarks and cautious deployment with human oversight before any institution relies on the models for production migration work.

The argument will sharpen once independent researchers attempt to replicate the MultiLangBench results or financial institutions publish case studies from actual COBOL migration projects. If OpenAI releases the benchmark publicly, outside evaluation will follow quickly and either confirm or complicate the headline numbers.