$830 billion disappeared in a week.

No hack. No exploit. No zero-day. No supply chain attack.

A feature update.

On January 13, 2026, Anthropic launched Claude Cowork — an AI agent that manages files, creates documents, extracts data, and automates office workflows. Fortune's headline said it "could threaten dozens of startups." Available for $20 a month on Pro. $200 on Max.

Seventeen days later, eleven specialized plugins. On February 5, Opus 4.6 with a one-million-token context window.

Three weeks. Three releases. Nearly one trillion dollars in market value evaporated.

It's not a bug. It's the roadmap.

The killchain

The sequence is surgical.

January 13. Claude Cowork. An agent that lives inside your workspace, reads your files, creates documents, extracts data from PDFs, organizes information. Everything that a hundred "AI assistant for work" startups were doing — integrated natively into Claude for $20 a month.

January 30. Eleven specialized plugins: Sales, Finance, Legal, HR. The legal plugin replaced functionality that cost thousands of dollars a month — compliance tracking, document review, due diligence that once required paralegals and a Thomson Reuters subscription.

Eleven knives. Eleven verticals. Eleven categories of startups that discovered, in a press release, that their product no longer had a reason to exist.

February 3-5. The market processes the information. What follows is called Software-mageddon.

February 5. Opus 4.6. One-million-token context window. The most capable model Anthropic has ever released to the public. Fortune's headline: "Anthropic's Claude triggered a trillion-dollar selloff. A new upgrade could make things worse."

It was right.

The body count

The S&P 500 Software Index fell roughly 4% in the first days of February. Sounds small. That is $830 billion.

The individual numbers are worse.

Thomson Reuters: -18%. The worst day in the company's history. The company that for decades was synonymous with "legal information" — Westlaw, regulatory databases, litigation analytics — lost nearly a fifth of its market value because a $20-a-month chatbot learned to read contracts.

LegalZoom: -20%. The promise of democratizing legal services died at the hands of something that democratized them faster.

FactSet: -10.5%. Specialized financial analytics. Terminals that cost tens of thousands a year. Claude's Finance plugin doesn't do everything FactSet does. Yet. The market priced in the "yet."

RELX (LexisNexis parent): -14%. Another legal fortress under siege.

iShares Software ETF: -5%. The entire index. Not a company. A category.

The total, including the extended selloff: nearly one trillion dollars. Not because of a scandal. Not because of a recession. Because of features Anthropic announced in a blog post.

The wrapper graveyard

There is a name for what is happening: the death of the AI wrapper.

The business model was simple. Take an LLM. Put a nice interface on it. Add some prompts. Charge a subscription. Celebrate your ARR. Pitch deck with a hockey stick. Series A. Team of 40. Office in SoMa.

The problem: you don't own the intelligence. You rent it.

Your product is a UI on top of an API on top of a model you don't control. Your "moat" is a prompt. Your differentiator is a system message that anyone can replicate in an afternoon. Your "proprietary technology" is an orchestration layer on top of a model whose creator can integrate it natively at any moment.

And just did.

99% of AI startups are predicted dead. Not because they fail. Because they work — until the platform absorbs their value. Not your model, not your product.

Sam Altman said it out loud: AI will be "quite harmful" to some software companies. A Google VP warned in February 2026 that LLM wrappers and AI aggregators have their "check engine light on."

The light is on. Nobody is braking.

The Jasper precedent

Nobody can say there was no warning.

Jasper AI was one of the first startups built on GPT. AI copywriting. Explosive growth. $1.5 billion valuation. The success story every pitch deck cited.

Then OpenAI launched ChatGPT.

Overnight, Jasper had a "formidable low-cost competitor." Not a competitor offering something similar — the same provider that powered Jasper was now offering the same capability directly to end consumers. Free.

Summer 2023: ARR forecast cut by over 30%. July 2023: layoffs. September 2023: CEO and CTO resign. Internal valuation cut 20%.

It didn't go bankrupt. Worse: it became irrelevant.

And the industry watched, nodded, and kept building wrappers.

The trap

There is an angle nobody talks about.

Claude Code Max costs $200 a month. For that price, a developer can manage a fleet of agents that build, deploy, and scale complex applications. A solopreneur can do the work of a team of 10. The barrier to starting a software company has been demolished.

But the barrier to understanding what you built has been demolished too.

Here is the hook: the entry price is irresistible. $200 a month for productivity that used to cost $20,000 in salaries. You build fast. You build a lot. You reach 100,000 lines of code in weeks.

Lines you didn't write. Lines you didn't carefully review. Lines that work — for now — on top of a model that can change tomorrow.

And then the question: if Anthropic raises the price to $800, what do you do? If the model changes and your code stops working, who fixes it? If the API goes down and you have an SLA with your client, who answers?

This is not hypothetical. It already happened.

Not you. You don't even fully understand what the model generated.

The dependency is not a side effect. It is the product.

The question nobody asks

The industry celebrates that a single person can build what once required a team of 50. They call it "the solopreneur unicorn." They sell it as democratization.

They don't ask what happens when the LLMs are gone.

Not "if." When.

Not as extinction — as change. A model gets deprecated. An API reprices. A company pivots. A regulatory regime bans certain outputs. The company selling you intelligence decides your vertical is no longer interesting, or worse, that it now competes with you.

What's left? A codebase nobody at your company can maintain without the model that wrote it. A tech stack whose logic lives in the head of an LLM that no longer exists. AI-generated documentation describing AI-generated code that was debugged by AI.

Turtles all the way down. Until the bottom turtle moves.

Three levels of blame

Level 1 — Anthropic. Built the weapon. Every feature release is a calculated shot. Cowork is not a product — it is a declaration: everything your startup does, we can do better, cheaper, and integrated. It is not accidental. It is the strategy.

Level 2 — The VCs and the founders. Funded and built companies whose only asset was an abstraction layer over someone else's model. They knew platform risk existed. They put it on slide 47 of the pitch deck, after the TAM and before the team slide. Nobody read it. Nobody wanted to read it.

Level 3 — The industry. Celebrated speed. Celebrated "move fast." Celebrated that anyone could build a SaaS in a weekend. Didn't ask what they were building on. Didn't ask who owned the foundation. Didn't ask what would happen when the owner of the foundation decided to build on top of it themselves.

They knew. Everyone knew.

$830 billion in a week. Thomson Reuters on its worst day in history. LegalZoom losing a fifth of its value because a $20 model learned to read contracts.

The question is not whether your startup survives Anthropic's next release.

The question is: what are you building that can't be replicated with a prompt?

And if the answer takes longer than five seconds, you already know what it is.