Bridging the Legacy Tech Divide
For decades, the global economy has relied on a backbone of older, low-resource programming languages. COBOL, for instance, still underpins a staggering 40% of online banking systems and processes the vast majority of in-person credit card transactions. Similarly, Fortran remains a cornerstone of meteorology and nuclear safety analysis. However, as the original developers retire and documentation fades, maintaining these systems has become a critical challenge. Addressing this, the latest AI research from OpenAI introduces new models capable of handling these niche languages with unprecedented accuracy and cost-efficiency.
The initiative is built on two primary pillars: the ability to construct functional software from scratch and the capacity to migrate ancient, mission-critical codebases into modern environments like Python or Rust without sacrificing functional parity. By focusing on these low-resource languages, the researchers aim to provide developers with the tools necessary to modernize infrastructure that was previously considered too risky or expensive to touch.
1. MultiLangBench: Building SQLite From Scratch
To rigorously test coding capabilities, the team developed MultiLangBench, a new benchmark that tasks an AI model with building a fully functional, SQLite-compatible database engine. The challenge requires the model to manage state, handle complex errors, organize a codebase, and implement new features while adhering to the strict requirements of 20 different languages, including Visual Basic, OCaml, and Zig.
The results are transformative. The new GPT-6 Astra model achieved an average test pass rate of 87.4% across the entire spectrum of languages, a massive leap over the 26.8% achieved by the previous GPT-5.6 Sol model. Perhaps most impressively, performance in notoriously difficult languages saw exponential growth: COBOL pass rates jumped from 10.9% to 72.1%, and Fortran performance surged from 21.4% to 83.5%. This benchmark is now available on GitHub, allowing engineers to audit how these models handle specific language quirks.
2. COBOL Migration: Preserving Business Logic
Modernization is not just about writing new code; it is about ensuring that existing business rules are preserved during a transition. The second major evaluation focuses on migrating COBOL programs to modern alternatives like C++ or Java. Because COBOL often handles precise financial rounding and specific transactional logic, even a slight error in translation can cause significant real-world issues. To combat this, the evaluation demands byte-for-byte output parity.
The model GPT-6.1 Sol has demonstrated remarkable efficacy here, reaching 86.6% accuracy at low reasoning levels, compared to the 5.4% accuracy of its predecessor. When pushed to higher reasoning efforts, the model hit 97.9% accuracy, essentially saturating the test suite. This represents a major turning point for enterprise software teams, as it offers a path toward reliable, automated migration that reduces the risk of logic drift in financial and industrial software.
Why It Matters
- Infrastructure Stability: By automating the maintenance of legacy systems, essential financial and safety-critical services can be updated with less risk of downtime.
- Cost Reduction: The new models deliver these high-accuracy results at significantly lower API costs—up to 78% cheaper than previous generations in some testing scenarios.
- Enterprise Scalability: The ability to migrate COBOL or Fortran codebases allows companies to move away from shrinking talent pools and toward modern, maintainable stacks.
