OpenAI Claims Major AI Gains in COBOL, Fortran and Visual Basic Coding

There’s a quiet crisis running underneath some of the world’s most important software. Banking systems, government databases, scientific computing platforms — many of them still depend on programming languages that most developers today have never touched. On 19 October 2026, OpenAI posted results suggesting its newest models are getting steeply better at working with exactly those languages.

The announcement focused on COBOL, Fortran and Visual Basic — three languages that have been around for decades and, in some cases, still quietly power essential services. OpenAI said its latest model advances improve software development and migration in these so-called low-resource programming languages, where training data is scarce and AI tools have historically struggled.

The Numbers Are Striking

OpenAI’s published results centre on two models: GPT-6 Astra and GPT-6.1 Sol. On a benchmark called MultiLangBench — which tests coding ability across 20 programming languages — GPT-6 Astra achieved an average pass rate of 87.4%. The previous model, GPT-5.6 Sol, managed 26.8% on the same test.

The language-specific jumps are even sharper. For COBOL, GPT-6 Astra went from 10.9% (GPT-5.6 Sol) to 72.1%. Fortran moved from 21.4% to 83.5%. Visual Basic climbed from 27.5% to 89.4%. Those are not incremental improvements — they represent a hefty change in what the model can apparently do with code that most modern developers find deeply unfamiliar.

Meanwhile, the migration results are where things get chiefly interesting. OpenAI ran a separate evaluation testing whether GPT-6.1 Sol could translate existing COBOL programmes into other languages — C++, Java, Kotlin, Python or Rust — while preserving the original behaviour. A task only counted as passing if the translated code produced identical outputs on hidden tests. At low reasoning effort, GPT-6.1 Sol hit 86.6% accuracy on that evaluation. GPT-5.6 Sol, by comparison, managed just 5.4%. Pushed to higher reasoning effort, GPT-6.1 Sol reached 97.9%.

OpenAI also reported that GPT-6.1 Sol cut estimated API costs by around 39% compared with GPT-5.6 Sol on the COBOL migration evaluation.

What Does This Actually Mean?

The context matters here. COBOL, Fortran and Visual Basic aren’t just museum pieces. COBOL in particular is embedded in financial and public-sector software that processes enormous volumes of transactions and records every day. The challenge with legacy systems isn’t just that they’re old — it’s that they’ve accumulated decades of embedded business rules, edge cases and institutional knowledge that often exists nowhere else. Rewriting them is expensive, risky and slow. Finding developers who know the languages is getting harder as the workforce that built those systems retires.

OpenAI said the model work is intended to help developers understand existing business logic, build new software and modernise systems while keeping existing behaviour intact.

But the caveats matter too. These are OpenAI’s own benchmarks, not independent assessments. The company has not published verified evidence of a live banking or public-sector system being migrated using these models. Benchmark performance in a controlled evaluation is a different thing from safely migrating a system that handles payroll, benefits payments or patient records.

Engineering teams and technology analysts have generally welcomed signs of progress in this area, while noting that production migration carries risks well beyond what any benchmark can capture — inaccurate code changes, confidential data handling, regulatory compliance and the sheer complexity of systems that have been modified over many years.

The Broader Picture for Regulated Industries

For organisations in financial services, healthcare, central government and local government, the bar for changing core software is high. In the UK, that means data-protection requirements under UK GDPR, operational-resilience rules from regulators like the Financial Conduct Authority and Prudential Regulation Authority, and procurement frameworks that govern how public bodies adopt new technology.

AI-assisted migration tools don’t remove those obligations. Any organisation considering using models like GPT-6.1 Sol to help migrate legacy code would still need security reviews, independent testing, rollback plans and human sign-off at every meaningful stage. The AI might reduce the effort involved in understanding old code or generating a first draft of a translation — but the responsibility for what goes into production stays with the people running the system.

That said, one practical benefit that technology teams have pointed to is the potential to address a shortage of specialists. Fewer developers each year have deep experience with COBOL or Fortran. If AI tools can help a team that doesn’t have that background to at least read, document and begin working with legacy code, that’s a real reduction in friction — even before any migration begins.

What This Means for Kent Residents

Kent’s public services — including Kent County Council, Medway Council and NHS Kent and Medway ICB — operate within national and sector-wide technology frameworks that, in some cases, connect to older systems. While no specific use of COBOL, Fortran or Visual Basic by Kent organisations has been independently verified, any improvements in AI-assisted legacy code tools that reduce the cost or complexity of modernising public-sector software could, over time, affect how efficiently local services are delivered. For now, OpenAI’s results are company-published benchmarks, and independent real-world evidence of large-scale operational benefits is uncertain.