OpenAI Says GPT-6 Has Got Much Better at COBOL and Other Old Programming Languages
OpenAI reports big gains on rarely used languages like COBOL and Fortran, including a jump on a COBOL migration test. All figures are the company's own.
OpenAI says its GPT-6 models are now far better at writing code in older, less common programming languages such as COBOL, Fortran and Visual Basic. According to The Decoder on October 20, these are “low-resource” languages, meaning there is much less public code to learn from than for Python or JavaScript. The reason this matters: The Decoder cites an IBM estimate that COBOL runs more than 40% of online banking systems and 95% of ATM transactions.
OpenAI backs the claim with two tests. The first, MultiLangBench, asks a model to build a database engine that works like SQLite from scratch, across 20 languages. At maximum reasoning effort, GPT-6 Astra averaged 87.4% of tests passed, and GPT-6.1 Sol came in at about 87%. The cheaper GPT-6 Luna scored 40.3%, only just ahead of GPT-5.6 Terra at about 38%. The second test covers COBOL migration, translating old COBOL into modern languages such as Java, Python or Rust. GPT-6.1 Sol reached 86.6% accuracy at low reasoning effort, compared with 5.4% for GPT-5.6 Sol, and 97.9% at higher effort.
Before getting excited, note the caveats. Every figure is OpenAI’s own, with no independent check. The two tests use different older models as the baseline, so the jumps cannot be compared directly. The coverage does not list the 20 languages, and it does not say how well the tests mirror real migrations, where code must be validated against live production systems.
What’s behind it: a “benchmark” is a standardised exam for AI models. Scores are useful for comparing, but a model that aces an exam may still stumble on a messy real-world bank system. Old languages are a hard case for AI because the training data is thin: most COBOL lives inside private company systems, not on public websites. The people who understand these systems are also retiring, so organisations have a real incentive to find help. That is why a model that can read and translate old code is commercially interesting.
What this means for you: if you are a student or a hobbyist, nothing changes, but it is a nice sign that AI help is spreading beyond the popular languages. If you work in a company with legacy software, this is a reason to run a small, well-supervised pilot, for example translating one low-risk module and checking the output against the original’s behaviour. Do not treat a benchmark percentage as a promise that your own system can be migrated safely.
Sources
Source: https://openai.com/index/low-resource-programming-languages/
Claude Designs Leaner Quantum Circuits, With a Computer Checking Every Step
Anthropic says Claude Mythos 5 found quantum circuit designs that need fewer qubits and gates than published ones across eight tasks. A built-in checker kept it honest.
