AivexaNewsSearch
AI news for builders and product teamsChecked every hour

OpenAI says GPT-6 is much better with legacy programming languages such as COBOL and Fortran

Collected Oct 20, 2026

OpenAI says it has improved its latest GPT-6 family of models for "low-resource" programming languages, the ones with far less public training code than mainstream languages like Python. It names COBOL, Fortran and Visual Basic among them, and argues these languages nonetheless support critical infrastructure around the world. To back that up, OpenAI cites IBM's estimate that COBOL runs more than 40 percent of online banking systems and 95 percent of all ATM transactions.

To track progress, OpenAI built two evaluations. The first, MultiLangBench, is published on GitHub and asks a model to construct a SQLite-compatible database engine from scratch, then measures test pass rates across 20 programming languages. With maximum reasoning effort, GPT-6 Astra reaches an average pass rate of 87.4 percent, compared with roughly 38 percent for GPT-5.6 Terra. GPT-6 Luna scores 40.3 percent, only slightly ahead of GPT-5.6 Terra's 37.9 percent but, according to OpenAI, at much lower cost. A chart accompanying the results shows GPT-6 Astra and GPT-6.1 Sol both landing near 87 percent on the average pass rate, with older models topping out around 38 percent.

The second evaluation targets COBOL migration, translating existing COBOL into C++, Java, Kotlin, Python or Rust. Here GPT-6.1 Sol reaches 86.6 percent accuracy at low reasoning effort, against 5.4 percent for GPT-5.6 Sol. Raising reasoning effort lifts accuracy to 97.9 percent. This suggests the biggest gains come when the model is allowed to spend more compute per task: the low-effort 86.6 percent is already a large step up, but the near-98 percent figure depends on settings that will be slower and more expensive per translation.

For developers and product teams, the practical signal is that legacy modernization, long a weak spot for code models because of scarce training data, is becoming a more credible use case. COBOL translation is the clearest example: a model that can reliably convert it to modern languages could lower the cost of re-platforming banking and insurance systems. A likely trade-off is that high accuracy on a benchmark does not guarantee correctness on a real codebase, so organizations would still need strong review and testing around any translated code.

These are OpenAI's own evaluations with its own reported numbers, so independent replication would be worth watching. The comparison also matters: Luna's near-flat result next to GPT-5.6 Terra suggests that not every model in the new family represents the same leap, meaning cost-sensitive deployments may see much less benefit than teams using the top-end models with high reasoning effort.

Why it matters: Teams maintaining COBOL, Fortran or Visual Basic systems are the clearest beneficiaries, since improved Model performance on low-resource languages makes automated migration and refactoring more plausible. In practice, developers should treat the reported pass rates as an encouraging benchmark rather than a guarantee, and plan for effort and cost trade-offs, especially where the highest accuracy requires maximum reasoning effort.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

OpenAI improved its latest models for low-resource programming languages like COBOL, Fortran, and Visual Basic. On the new MultiLangBench, GPT-6 Astra hits an 87.4 percent test pass rate across 20 languages, up from about 38 percent for GPT-5.6 Terra. For COBOL migration specifically, GPT-6.1 Sol reaches up to 97.9 percent accuracy, a result that could matter for modernizing critical financial infrastructure. The article OpenAI says GPT-6 is much better with legacy programming languages such as COBOL and Fortran ap