> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI Built an AI That Reads 65-Year-Old Mainframe Code Nobody Understands Anymore
- URL: https://wire.fourthweb.ai/openai-built-an-ai-that-reads-65-year-old-mainframe-code-nobody-understands-anymore/
- Published: 2026-10-20T00:00:54.000Z
- Updated: 2026-10-20T00:00:57.000Z
- Description: The next trillion dollars in productivity is locked inside mainframes running code from 1959, and OpenAI just built the key.
- Author: Travis Wright
- Tags: AI Agent Economy, AI Agents, OpenAI, Funding Rounds

**The next trillion dollars in productivity is locked inside mainframes running code from 1959, and** [**OpenAI**](https://wire.fourthweb.ai/tag/openai/) **just built the key.**

### The Summary

- [OpenAI released new evaluations and benchmarks](https://openai.com/index/low-resource-programming-languages?ref=wire.fourthweb.ai) for ChatGPT's performance in low-resource programming languages including COBOL, SQLite, and specialized software engineering tasks
- This targets the massive enterprise legacy code problem: billions of lines of mission-critical code that almost nobody knows how to maintain anymore
- The move signals [AI agents](https://wire.fourthweb.ai/tag/ai-agents/) are shifting from writing fresh Python to understanding and migrating decades-old systems that run banking, insurance, and government infrastructure

### The Signal

Low-resource programming languages are not exotic. They are the foundation of critical infrastructure that processes your paycheck, routes your flight, and tracks your mortgage. [OpenAI's new benchmarks](https://openai.com/index/low-resource-programming-languages?ref=wire.fourthweb.ai) for COBOL, SQLite, and domain-specific software engineering represent something more important than better code completion. They represent the first serious attempt to give AI agents fluency in the languages that actually run the world economy.

COBOL alone powers an estimated 220 billion lines of code in production today. The programmers who wrote it are retiring or dead. The cost to migrate those systems has been prohibitive enough that banks and governments have simply accepted the risk of running 40-year-old code nobody fully understands. That math just changed.

> "The next trillion dollars in productivity is locked inside mainframes running code from 1959, and OpenAI just built the key."

What makes a programming language "low-resource" in the AI context:

- Minimal training data compared to Python, JavaScript, or Java
- Sparse documentation and examples available online
- Few active developers creating new code samples
- Limited Stack Overflow threads and GitHub repositories to learn from

[OpenAI's new evaluations](https://openai.com/index/low-resource-programming-languages?ref=wire.fourthweb.ai) measure ChatGPT's ability to understand, debug, and migrate code in these languages. This is not about writing net-new COBOL. It is about reading legacy COBOL, understanding what it does, and either fixing it or translating it into something modern. The SQLite work matters for different reasons: it is a widely deployed embedded database that lives inside applications, often with complex, hand-optimized queries that are hard to refactor.

The timing connects to a broader shift. AI coding assistants started with autocomplete for popular languages. Then moved to generating full functions. Now they are targeting the unsexy but valuable work of understanding and modernizing legacy systems. That is where the money is. Fortune 500 companies spend billions annually just maintaining old code, not improving it.

### The Implication

If you work in enterprise IT, this is your roadmap for the next three years. AI agents that can reliably read and migrate COBOL mean the multi-decade modernization backlog becomes tractable. Companies that have deferred mainframe migrations due to cost and risk now have a third option beyond "keep paying maintenance" and "rewrite from scratch." They can use AI agents to incrementally translate, test, and verify migrations at a fraction of the cost.

For developers, the skill shift is clear. Understanding legacy systems and guiding AI through migrations becomes more valuable than hand-coding in those languages. The question is not whether AI will handle COBOL. The question is who will prompt it correctly, verify its work, and manage the risk when it translates code that processes millions of transactions per day.

### Sources

[OpenAI Blog](https://openai.com/index/low-resource-programming-languages?ref=wire.fourthweb.ai)