OpenAI says its newest artificial intelligence models are markedly better at rewriting ageing software written in older languages into modern code — a claim aimed at organisations stuck with systems they dare not break.
On Monday the company pointed to what it called substantial improvements in migrating legacy programmes, including COBOL, into languages such as C++, Java, Kotlin, Python and Rust. It stressed that a migration counts as successful only if all tests pass, meaning the new code must behave like the old.
What OpenAI reported
In OpenAI’s own evaluations, GPT-6.1 Sol preserved software behaviour with 87 percent accuracy at low reasoning effort — a measure of how much computational analysis the model applies before answering — and 98 percent when that effort was raised. At low effort, GPT-6 Astra scored about 86 percent, GPT-6 Sol 21 percent and GPT-5.6 Sol only 5 percent.
The company also reported stronger results when models build database software in several languages, including COBOL. OpenAI framed the work as progress toward more accurate and affordable code migration for teams modernising software their businesses depend on.
COBOL — short for common business-oriented language — dates to 1959 and still underpins large parts of banking and government computing. 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, and it runs in many public systems from Social Security to the US Internal Revenue Service. One rough estimate puts the yearly cost of COBOL inefficiencies to the US economy at around $105 billion. Fortran, another older language, remains in scientific and security-related computing.
Rival pitch and market reaction
Anthropic, another AI firm, published a code modernisation playbook in February and promoted its Claude Code agent as a way to move COBOL into Java or Python. It offered no performance benchmarks for those migrations but said the economics of COBOL modernisation had shifted. Shares in IBM, which earns part of its revenue helping clients maintain legacy systems, fell sharply after that announcement.
Why experts still urge caution
Researchers who study software evolution said AI tools have improved on legacy code in recent years, yet remain far from safely replacing critical systems. OpenAI spokesperson Oscar Haines said legacy code is still far from a solved problem and that the company is taking feedback from developers and enterprises while exploring better tests and deeper language understanding.
Vadim Zaytsev, an associate professor of software evolution at the University of Twente in the Netherlands, argued that rewriting code is not the same as replacing a system. He said large COBOL estates can run to hundreds of millions of lines — the biggest he had seen had 250 million lines — and sit alongside mixed languages, varied data formats, integrations and organisational expectations. In his view, AI is not yet playing a major role in replacing or migrating legacy COBOL systems.
Arie van Deursen, a software engineering professor at Delft University of Technology, called OpenAI’s published evaluation limited, saying the migration task was thinly described and that the benchmark resembled relatively simple, well-defined work of the kind models may already have seen in training. He stressed that maintaining ageing software is far easier than modernising it, and that the company’s test tasks sit a long way from real-world modernisation.
Timeline: what has happened so far
- 1959: COBOL is created as a common business-oriented programming language.
- February: Anthropic releases a code modernisation playbook and pitches Claude Code for migrating COBOL to languages such as Java or Python; IBM shares fall after the announcement.
- Monday: OpenAI reports higher behaviour-preservation scores for migrating COBOL and other legacy code with its latest models, including GPT-6.1 Sol at up to 98 percent with higher reasoning effort.
Aasan alfaaz mein
OpenAI ka kehna hai ke us ke naye AI models purane COBOL software ko nayi languages mein badalne mein pehle se bohot behtar ho gaye hain. Banking aur ATM jese systems ab bhi bohot had tak isi purani language par chalte hain, is liye yeh khabar important hai. Lekin experts kehte hain ke sirf code likh dena system badalne ke barabar nahi — abhi asli, bohot bari systems ko AI se safe replace karna door ki baat hai. Matlab: ummeed hai, lekin aam logon ke banking systems mein foran koi bada tabdeeli prove nahi hui.
