On February 23, 2026, a single Anthropic blog post wiped out roughly $30 billion of IBM's market cap in one trading session — the company's worst single-day stock drop in 26 years. The trigger was not a scandal, a missed earnings report, or a lawsuit. It was a claim that Claude Code, Anthropic's AI coding assistant, could automate the most expensive phases of COBOL modernization: the discovery, analysis, and documentation work that has historically required large teams of consultants and years of effort.
This article breaks down exactly what happened, what both sides are arguing, and what it actually means for enterprises running legacy mainframe systems.
What Is COBOL, and Why Does It Still Matter?
COBOL — Common Business-Oriented Language — was created in 1959. By any reasonable measure, it should be a museum exhibit. Instead, it is the backbone of global finance. An estimated 95% of ATM transactions in the US run on COBOL, and hundreds of billions of lines of COBOL run in production every day, powering critical systems in finance, airlines, and government.
The problem is a slow-motion talent crisis. The developers who built these systems retired years ago, and the institutional knowledge they carried left with them. Production code has been modified repeatedly over decades, but the documentation hasn't kept up. COBOL is taught at only a handful of universities, and finding engineers who can read it gets harder every quarter.
This combination — mission-critical code, no documentation, vanishing expertise, and sky-high consulting costs — is what created IBM's highly profitable COBOL modernization business in the first place.
What Anthropic Actually Claimed
On February 23, Anthropic published a technical blog post positioning Claude Code directly against this problem. The post describes a structured methodology: Claude Code reads the full COBOL codebase and maps dependencies, identifies program entry points, traces execution paths, and surfaces implicit couplings through shared data structures and file operations that static analysis tools miss. Workflow documentation follows automatically, generating diagrams of processing pipelines that exist only in the code itself. Risk assessment then identifies high-coupling modules, isolated components ready for early migration, and areas of accumulated technical debt.
The core claim was economic: legacy code modernization stalled for years because understanding legacy code cost more than rewriting it. AI flips that equation. Anthropic said teams could now modernize COBOL codebases in quarters, not years. It also published a Code Modernization Playbook alongside the post.
The Market Reaction
IBM shares recorded their worst single-day drop in more than 25 years, plunging 13% after Anthropic published the post. But the damage was not limited to IBM. Accenture and Cognizant also declined following the news — a sign that investors were looking at the entire consulting model around legacy modernization, not just IBM's mainframe hardware business.
The pattern is becoming familiar. Just the week before, cybersecurity stocks sold off sharply after Anthropic announced Claude Code Security, a tool that scans codebases for vulnerabilities. Each new AI capability announcement triggers a reassessment of which existing revenue streams might be compressed, and the market prices in fear immediately.
IBM's Response: Translation Is Not Modernization
IBM did not stay quiet. Rob Thomas, Senior Vice President and Chief Commercial Officer at IBM, published a response arguing there is a clear difference between translating code and modernizing a platform: "Translation captures almost none of the actual complexity. Decades of hardware-software integration cannot be replicated by moving code."
Thomas made two pointed arguments:
1. The platform, not the language, is the value. Thomas compared COBOL on IBM Z to iOS on iPhone: "Someone could build an alternative, but it is unlikely to displace a billion iPhones. The performance derives from tight coupling of software and hardware, processor-level acceleration, I/O subsystem optimization, and decades of performance tuning."
2. COBOL is not mostly a mainframe problem. Thomas noted that nearly 40% of COBOL doesn't even run on mainframes — it runs on Windows, Linux, and other distributed platforms. Much of what the market was framing as an IBM mainframe story was actually a distributed systems problem.
Industry analysts echoed the pushback. John McKenny, SVP and general manager for intelligent Z optimization and transformation at BMC Software, said plainly: "Translating code itself isn't modernization."
The Real Scope of COBOL Modernization
Here is where the gap between Anthropic's claim and IBM's rebuttal comes into focus. Both are, in different ways, correct.
Two things can be true at once: AI tools, including Claude Code, are genuinely useful for understanding legacy COBOL codebases. And the hard part isn't reading or translating the code. It's data architecture redesign, transaction processing integrity, runtime replacement, and proving the new system does exactly what the old one did.
Successful COBOL modernization programs require business scoping, technical assessment, data migration planning, behavioral equivalence validation, observability, and organizational change management, in addition to code translation.
Where Claude Code has genuine impact is the discovery phase — the work that has traditionally consumed the most consulting budget and time before a single line of code is actually changed.
| Phase | What It Involves | Claude Code's Role |
|---|---|---|
| Discovery & Analysis | Mapping dependencies, documenting workflows, identifying risks | High — can automate significantly |
| Code Translation | Converting COBOL to modern languages (e.g., Java) | Partial — useful but requires oversight |
| Data Migration | Moving and validating data structures | Low — mostly human and specialized tooling |
| Behavioral Equivalence | Proving new system matches old system behavior | Low — requires domain expertise |
| Compliance & Security | Meeting regulatory requirements (FedRAMP, HIPAA, PCI) | Low — embedded tooling needed |
| Organizational Change | Training teams, updating procedures, managing risk | None — entirely human |
IBM's Own AI Strategy: Project Bob and watsonx Code Assistant
A detail largely missed in the market panic: IBM has been building AI tools for COBOL modernization for years.
IBM launched watsonx Code Assistant for Z in August 2023. It uses a 20-billion-parameter model trained on COBOL-Java pairs to help developers selectively refactor COBOL into Java — including application discovery, dependency mapping, automated refactoring, and validation. CEO Arvind Krishna said in July 2025 that the tool "has got very wide adoption."
More recently, IBM's Project Bob — an AI-first IDE built on VS Code — runs a multi-model architecture that routes Claude, Mistral, Meta's Llama, and IBM's own models to tasks based on accuracy, latency, and cost. IBM i CTO Steve Will confirmed Project Bob will replace watsonx Code Assistant for i, covering RPG, CL, SQL, COBOL, Java, and Python on IBM i at launch.
Project Bob also embeds FedRAMP, HIPAA, and PCI compliance context and inline security scanning directly in the IDE — which matters for regulated industries where the audit trail starts at the first line of changed code.
Notably, Project Bob actually uses Claude as one of its underlying models. This is not a clean Anthropic-vs-IBM battle. It is more complicated than that.
Real-World Results: What Enterprises Are Seeing
Royal Bank of Canada has used IBM's watsonx Code Assistant for Z to map dependencies and build modernization blueprints for core applications. The National Organisation for Social Insurance reported a 94% reduction in time to analyse legacy COBOL code using the same tool — cutting an eight-hour task to roughly 30 minutes.
These results support Anthropic's core economic argument: AI genuinely compresses the discovery and analysis phases. They do not, however, suggest that the rest of modernization has been solved.
Was the Market Selloff Justified?
The honest answer is: partially.
| Concern | Verdict |
|---|---|
| AI threatens IBM's consulting discovery revenue | Legitimate — this phase is genuinely automatable |
| AI can replace full COBOL modernization programs | Overstated — complex integration work remains human |
| IBM has no AI response | False — watsonx, Project Bob, and multi-model approaches already exist |
| IBM mainframe hardware is immediately at risk | Unclear — IBM reported its highest mainframe revenue in 20 years just before the selloff |
| Accenture/Cognizant consulting revenue is threatened | Legitimate concern for the discovery/analysis segment specifically |
Evercore ISI analyst Amit Daryanani noted that IBM has already provided customers with several modernization options and that clients who had the option to migrate from the mainframe have been sticking with the platform.
IBM's strength remains its deep integration in highly regulated industries — banking, healthcare, government — where "moving fast and breaking things" is not an option.
What This Means for Enterprises Running COBOL
If you are a CTO, architect, or IT leader responsible for legacy systems, here is the practical takeaway:
Claude Code is a legitimate tool for the discovery phase. If your modernization program has stalled because analysis and documentation costs are prohibitive, AI tools — including Claude Code — have genuinely changed the economics of that specific phase. This is real, not hype.
It does not replace the full modernization stack. As Steven Perva, expert mainframe innovation engineer at Ensono, put it: understanding that you "most likely could teach these LLMs about these other systems" is necessary, but "when we're talking critical enterprise infrastructure, 99.99% uptime isn't good enough" as an assumption. Data migration, behavioral equivalence validation, compliance controls, and organizational change remain human-intensive work.
The best approach probably uses both. The best approach probably isn't Anthropic or IBM — it's both, along with the systems integrators, testing tools, and domain experts who understand why these systems were built the way they were.
The Bigger Pattern
This episode is not really about COBOL. It is about a recurring dynamic in 2025 and 2026: AI tools that automate the expensive, time-consuming analysis phase of a professional service engagement, without yet replacing the execution, validation, and governance phases that follow.
The same pattern played out in cybersecurity when Claude Code Security launched. It is playing out in legal document review, financial analysis, and software auditing. In each case, the market prices in total disruption; the reality is more targeted disruption of the highest-margin, lowest-differentiation work — the billable hours spent getting to "we now understand the problem."
That is still significant. But it is different from replacing the entire professional services model.
Bottom Line
Anthropic's Claude Code can genuinely accelerate COBOL modernization — specifically the dependency mapping, workflow documentation, and risk identification phases that have historically made the discovery process prohibitively expensive. That is a real shift in the economics of legacy modernization.
IBM's counter-argument is also valid: the mainframe's value is not the COBOL code. It is the platform, the hardware-software integration, the compliance infrastructure, and the decades of reliability. Code translation is not platform replacement.
The market reaction — a 13% single-day drop for IBM — likely overshot. But the underlying question it raised is the right one: as AI systematically compresses the cost of the most expensive phases of professional services work, which consulting revenue streams survive, and which do not?
For COBOL modernization, the answer in 2026 is that the discovery phase is under genuine pressure. Everything that comes after it is not.