OpenAI Improves ChatGPT for COBOL, SQLite and Low-Resource Code

OpenAI has detailed significant improvements to ChatGPT's capabilities for software engineering in low-resource programming languages. According to an announcement from the company, this effort includes new evaluations focused on tasks involving SQLite and the migration of legacy COBOL code, aiming to bridge a persistent performance gap for languages that are less represented in standard AI training data.

For creators and developers working with niche or legacy systems, this development matters. It signals a move beyond AI coding assistants that excel only in popular, modern languages like Python and JavaScript, toward tools that can understand and help modernize the critical, often ancient, code that still powers vast segments of finance, government, and enterprise infrastructure.

What are low-resource programming languages?

In the context of AI model training, a low-resource programming language is one for which there is a relative scarcity of publicly available, high-quality code examples and documentation. This scarcity limits the model's ability to learn the syntax, patterns, and idioms of the language effectively. While Python and JavaScript are considered high-resource, languages like COBOL, Fortran, Ada, and even specialized dialects or older versions of more common languages often fall into the low-resource category. The challenge isn't just the language itself, but also the specific contexts—like mainframe environments or proprietary databases—in which they are used.

What specific improvements did OpenAI announce?

OpenAI's announcement, published on October 19, 2026, centers on new evaluation benchmarks. The company stated it has developed and applied these tests to measure ChatGPT's performance in two key, practical areas. The first is working with SQLite, a widely used database engine where precise schema understanding and query generation are crucial. The second, and more notable, is assisting with COBOL migration projects. COBOL is a decades-old language that remains foundational to countless legacy banking and administrative systems, yet there is a severe shortage of developers who can maintain it. The new evaluations are designed to track progress in helping developers understand, refactor, and translate this critical legacy code.

Why is better COBOL and SQLite support important?

The importance is twofold: risk and cost. Millions of lines of COBOL code run global financial transactions and government services. Every year, the pool of developers who can maintain this code shrinks, creating a massive institutional risk. AI that can reliably help analyze and migrate this code is a strategic tool for business continuity. For SQLite, which is embedded in nearly every mobile application and many desktop programs, improved AI assistance means fewer database-related bugs and more efficient data handling. Together, these improvements target real-world pain points where developer time is expensive and errors are costly.

How does this relate to AI coding in Grafy?

This advancement aligns directly with the practical, multi-model coding environment offered in Grafy Code. Grafy Code is an AI coding studio where developers can work across different models and languages in a single workspace. The improvements highlighted by OpenAI suggest that the underlying models available through services like OpenRouter—which are accessible within Grafy—are becoming more capable for specialized tasks. For a developer tackling a COBOL migration or optimizing a complex SQLite integration, they could use Grafy Code to prompt multiple AI models side-by-side, comparing their suggestions for legacy code or generating test cases. The platform's ability to run code and install packages in a sandboxed environment also allows for practical validation of AI-generated snippets for these languages.

For those researching this trend, Grafy Reader offers a way to analyze the original OpenAI announcement and related technical papers. By uploading the source document to Reader, you can have an AI quickly summarize the key points, extract the evaluation methodologies, and answer specific questions about the benchmarks, all while providing citations back to the source text. This is invaluable for separating the concrete claims from the broader narrative.

What does this mean for the future of AI-assisted development?

The move signifies a maturation of AI coding tools from general-purpose assistants to specialized partners for specific engineering challenges. It acknowledges that the future of software isn't just about writing new code in trendy languages, but also about stewarding and transforming the enormous existing codebase that the digital world is built upon. This development follows a broader industry trend, as seen with other specialized models like Reflection's Beam for coding and agents, and Nvidia's Nemotron models for enterprise use. As these evaluations become standard, we can expect more AI providers to compete on performance in these niche but critical domains, ultimately giving developers more powerful tools for system modernization.

Frequently asked questions

What is a low-resource programming language?

It's a programming language with relatively little high-quality, publicly available code for AI models to learn from. This includes legacy languages like COBOL and Fortran, as well as specialized DSLs. AI assistants have traditionally performed worse on them compared to high-resource languages like Python.

How can I try this with COBOL code in Grafy?

Use Grafy Code, the AI coding studio. You can paste a snippet of legacy COBOL code and prompt connected AI models to explain it, suggest refactors, or even generate migration outlines to a modern language. Compare outputs from different models to assess the best approach.

Will this make COBOL programmers obsolete?

No, it's the opposite. These tools augment and empower the remaining COBOL experts. The AI handles tedious analysis and translation drafts, but a human engineer's domain knowledge is essential for validating logic, understanding business rules, and ensuring the migrated system works correctly.

---

Source: OpenAI News — Advancing coding in low‑resource programming languages