AI and COBOL Modernization: Why IBM’s Stock Fell After Anthropic’s Announcement

IBM shares fell 13.2% on February 23, 2026, after Anthropic described how Claude Code could help modernize COBOL systems. Reuters reported that it was IBM’s steepest daily decline since October 2000.

The reaction reflected concern that AI could reduce demand for some legacy-system consulting work. It did not establish that mainframes had become obsolete, or that an AI-generated translation could safely replace a production system without further engineering.

Why COBOL modernization attracts attention

COBOL remains part of long-running business systems, including financial and government applications. In a modernization project, understanding existing behavior can be as important as writing replacement code. Business rules may be distributed across programs, data definitions, scheduled jobs, and integrations.

Replacing a language therefore involves more than making new code compile. The team needs to know what the old system does, which behaviors must remain, and how to detect a change that would affect users or business operations.

What Anthropic says AI can help with

Anthropic’s modernization playbook presents agentic coding as a way to assist code analysis, migration, documentation, and test generation. Its claims should be understood as a proposed workflow, not a universal delivery schedule for every COBOL system.

Claude Code can work with project files and tools, as illustrated in TipsMake’s multi-file coding guide. Access to more code can support analysis, but it does not guarantee that the agent has found every dependency or correctly understood an undocumented rule.

IBM’s response: the platform still matters

In a February 23 response, IBM executive Rob Thomas distinguished language translation from platform modernization. IBM argues that mainframe value also comes from the surrounding operating environment, transaction processing, security, and operational resilience.

That is IBM’s position in a debate affecting its business. It nevertheless identifies a practical distinction: translating a program does not automatically migrate its data, integrations, runtime behavior, and operational controls.

What a migration team still needs to prove

A contained pilot with measurable checks is more informative than a promise that a whole migration will take quarters rather than years. Compare the old and new systems’ behavior before expanding the scope.

When choosing an assistant, distinguish code suggestions from tools that can inspect and change a project; the Claude Code and GitHub Copilot comparison discusses that workflow difference. The broader question is which parts of modernization become cheaper, and which still require specialist judgment. A single trading day cannot settle it.