OpenAI reported substantial improvements on October 19th in its latest AI models' ability to migrate aging legacy code written in older programming languages like COBOL into modern alternatives, potentially helping organizations maintain older systems, according to The Verge.
Key takeaways
- OpenAI highlighted progress in moving COBOL code to languages like C++, Java, Kotlin, Python, and Rust.
- In tests, the GPT-6.1 Sol model achieved up to 98 percent behavior preservation accuracy at high reasoning effort.
- Anthropic also pitched its Claude Code agent earlier in February as a tool for COBOL modernization.
- Experts consulted by The Verge warn that rewriting code is very different from replacing entire complex systems.
Can AI help solve the COBOL code crisis?
Banks, airlines, the IRS, and government infrastructure around the world rely on antique computer code that nobody dares touch. Short for common business-oriented language, COBOL came into being in 1959. According to IBM, it forms the foundation for more than 40 percent of online banking systems, supports 80 percent of in-person credit card transactions, and handles 95 percent of ATM transactions. One rough estimate puts the annual cost of COBOL inefficiencies to the US GDP at around $105 billion.
Replacing these systems is difficult because they are too important to risk breaking, and many original programmers have retired or died. OpenAI reported promising results in migrating COBOL code without changing original software behavior. In OpenAI tests, GPT-6.1 Sol achieved 87 percent behavior preservation accuracy at low reasoning effort, climbing to 98 percent at higher reasoning. By comparison, GPT-6 Astra scored around 86 percent, GPT-6 Sol scored 21 percent, and GPT-5.6 Sol scored 5 percent.
How does COBOL migration compare to other legacy efforts?
Anthropic sees a similar opportunity in legacy systems. The company published a code modernization playbook in February and pitched its Claude Code agent for moving COBOL into Java or Python. Shares in IBM, which helps companies maintain legacy systems, nosedived following that announcement.
Despite these updates, experts interviewed by The Verge remain cautious. Vadim Zaytsev, an associate professor at the University of Twente, noted that rewriting code is not the same as replacing a system. Systems involve multiple languages, varied data formats, and technical integrations. Arie van Deursen, a software engineering professor at Delft University of Technology, described OpenAI's blog post as superficial and noted that tasks used for assessments were relatively simple compared to real-world infrastructure.
What it means for organizations relying on antique software
Companies and government agencies depending on decades-old infrastructure face a long road before automated tools can fully take over. OpenAI spokesperson Oscar Haines acknowledged that legacy software is still far from a solved problem, noting the company is listening to developer feedback and working to improve model testing. For now, the technology can assist with specific rewriting tasks, but enterprise tech teams must still carefully evaluate whether automated outputs preserve the exact behavior needed for critical public services and financial networks.
Frequently asked questions
COBOL stands for common business-oriented language and came into being in 1959. According to IBM, it supports more than 40 percent of online banking systems, 80 percent of in-person credit card transactions, and 95 percent of ATM transactions, alongside government systems like the IRS.
On October 19th, OpenAI touted substantial improvements in its latest AI models' ability to migrate aging software written in older programming languages like COBOL into modern code without changing original software behavior.
In OpenAI tests, the GPT-6.1 Sol model achieved 87 percent behavior preservation accuracy at low reasoning effort, climbing to 98 percent at higher reasoning. GPT-6 Astra scored around 86 percent, GPT-6 Sol scored 21 percent, and GPT-5.6 Sol scored 5 percent.
In February, Anthropic published a step-by-step code modernization playbook and pitched its Claude Code agent as a way to migrate COBOL code into modern languages like Java or Python.
Experts like Vadim Zaytsev and Arie van Deursen explain that rewriting code is not the same as replacing a system. Real systems involve multiple languages, varied data formats, and complex technical integrations that automated benchmarks do not fully capture.
Based on reporting by The Verge. Read the original article
