What happened in this Tech story?

OpenAI reported major gains in behavior-preserving COBOL migration, but academic experts warned that its benchmark leaves the harder work of decomposing and validating critical legacy systems largely unresolved in practice.

What is a key fact in this story? (1)

**OpenAI** said on October 19 that its latest models made “substantial improvements” in migrating aging software, reporting behavior-preserving COBOL conversions to C++, Java, Kotlin, Python and Rust; it defined success as every test passing.

What is a key fact in this story? (2)

**GPT-6.1 Sol** reached 87% behavior-preservation accuracy at low reasoning effort and 98% at high effort in OpenAI’s tests, versus 86% for GPT-6 Astra, 21% for GPT-6 Sol and 5% for GPT-5.6 Sol at low effort.

📎 SkimNews has covered OpenAI 715+ times · see the file →

Can AI help solve the COBOL code crisis? — SkimNews

By The Verge · Summarized & edited by Kapil Samant · 2026-10-26

Get the Tech newsletter

Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.

OpenAI said on October 19 that its latest models made “substantial improvements” in migrating aging software, reporting behavior-preserving COBOL conversions to C++, Java, Kotlin, Python and Rust; it defined success as every test passing.
GPT-6.1 Sol reached 87% behavior-preservation accuracy at low reasoning effort and 98% at high effort in OpenAI’s tests, versus 86% for GPT-6 Astra, 21% for GPT-6 Sol and 5% for GPT-5.6 Sol at low effort.
Anthropic published a February code-modernization playbook promoting Claude Code for COBOL migration into Java or Python, but supplied no performance benchmarks; IBM shares nosedived after the announcement.
IBM estimates COBOL supports more than 40% of online banking systems, 80% of in-person credit-card transactions and 95% of ATM transactions, while one rough estimate puts its annual inefficiencies at around $105 billion to the U.S. GDP.
Vadim Zaytsev said the largest COBOL system he had seen held 250 million lines and that AI is not playing a major role in migrating such systems, whose code, data, integrations and dependencies must be divided into replaceable units.
Arie van Deursen called OpenAI’s blog “superficial,” saying its relatively simple, well-defined benchmark was “lightyears” from modernization because it did not test breaking complex legacy systems into individually replaceable units.
Jacopo Mauro said formal verification could automatically check a translation better than a human, but adapting it to legacy code would require a “huge amount of work”; human verification remains essential.

Why it matters: In critical systems such as those used by the IRS and banks, even minor behavioral differences can be unacceptable, so GPT-6.1 Sol’s 87% low-reasoning preservation score is not a safe replacement standard. With COBOL handling 95% of ATM transactions, the practical work remains decomposition, testing, formal verification and expert review.