Disclosure: I run Lago, an open-source billing engine. I'm on Claude Code and ChatGPT daily. I've used Cowork since its January launch.
Cyber psychosis
Last week I was at a CEO gathering organized by our lead investor. Series A to Series D founders. Different industries, different stages, different problems. Every single one was on Claude Code.
Not "evaluating." Not "piloting." Using it. Daily. The conversation wasn't whether to use AI tools. It was how to implement Claude across the organization. Governance, use cases, driving adoption.
We dropped Cursor. So did most teams around us. A few YC batches ago, everyone was on Windsurf. Before that, Cursor. The rotation is fast. This time feels different. Teams aren't moving to the next IDE wrapper. They're moving to the model provider directly. Wrappers are fragile. When the underlying model is the product, you go to the source.
Garry Tan, YC's CEO, [open-sourced his entire Claude Code setup](https://x.com/garrytan/status/2032014570118922347) in March. 20,000 GitHub stars in 48 hours. He claims 10,000 lines of code and 100 PRs per week. He sleeps four hours a night. His words [at SXSW](https://techcrunch.com/2026/03/17/why-garry-tans-claude-code-setup-has-gotten-so-much-love-and-hate/): "I have cyber psychosis, but I think a third of the CEOs that I know have it as well." Karpathy [posted in December](https://x.com/karpathy/status/2015883857489522876) he'd gone from 80% manual coding to 80% agent coding in weeks. Called it "the biggest change in two decades of programming."
That's the state of things. Anthropic looks like it killed the game.
It didn't.
The real game: smartest vs. most embedded
A quick distinction first, because people confuse the two: Claude Code is a terminal tool for developers. Cowork is a desktop app for everyone else. Same models underneath. Different surfaces. That line is already blurring. The "coding tool vs. workflow tool" framing might not survive 2026.
In January, Anthropic shipped Cowork as a research preview. By March, Microsoft built Copilot Cowork on Anthropic's models. Three months. Enterprise teams went from zero to running multi-step agents across Gmail, Notion, Google Drive, and Slack. More than half of usage comes from non-technical roles. Marketing, ops, HR, finance.
The speed is real. But speed of adoption isn't the same as winning. The question isn't which AI is smartest. It's which AI becomes most embedded. Distribution, integrations, switching costs. Not benchmarks.
And on that question, Anthropic has serious competition.
The Codex counterargument
Codex has 2 million weekly active users. Cisco, Nvidia, Ramp, Rakuten deployed it across engineering teams. Sam Altman [calls it](https://x.com/sama/status/1992385857489522876) "the most loved internal product we've ever had." Peter Steinberger built OpenClaw entirely with Codex, liked it so much he [joined OpenAI](https://techcrunch.com/2026/02/15/openclaw-creator-peter-steinberger-joins-openai/) in February.
"Just a coding tool" is what people say now. OpenAI is explicit about what comes next: Codex as a general enterprise agent platform. [Fortune described it](https://fortune.com/2026/03/04/openai-codex-growth-enterprise-ai-agents/) as "at the heart of OpenAI's plans to sell AI agents to enterprises." And ChatGPT can already connect to email, calendar, and other work tools. Most people don't know. If OpenAI makes those integrations a first-class product instead of a buried feature, that changes the math overnight.
The engineering-first wedge is strategically smart. Engineers are the hardest audience to win. If Codex earns their trust and expands into cross-functional workflows, that's a bottom-up motion that's very hard to reverse. Codex in 2027 could look like Cowork today. With a developer-trust moat Anthropic doesn't have.
Three other moats
Google: already in your stack. Gemini Enterprise. Workspace Studio. Non-technical users building custom agents inside Google Workspace. No code. $21/seat/month as part of existing plans. "It's already included in your Workspace plan" is a procurement argument that no technical superiority can counter. That's how Google Docs ate Microsoft's lunch in education and SMB. Same playbook. Different decade.
Meta: the Linux play. Meta isn't selling enterprise seats. It's giving away the models. Llama 4 is open-source, multimodal, and good enough. Llama Stack is becoming the deployment standard for companies that want AI on their own infrastructure. IBM, Red Hat, Dell are building on top. 3 billion users on WhatsApp and Messenger is distribution nobody else has.
Mistral: sovereignty and speed. Mistral is the one people keep underestimating. Global AI company, not a regional play. Mistral Medium 3 is competitive with frontier models. Le Chat Enterprise shipped with self-hosting, connectors for SharePoint, Google Drive, Gmail, and a built-in agent builder. Where Mistral has an edge none of the US companies can match: data sovereignty. Self-hosted deployment. Infrastructure that stays in your jurisdiction. That's not just an EU story. Banking, healthcare, defense, government, any regulated industry in any region has data residency requirements. Mistral can sell to all of them. Anthropic and OpenAI can too, and have, but they have many fish to fry already. Mistral's entire product is built around this. That focus matters. That's a significant chunk of the Fortune 500.
What Anthropic actually got right
The moat isn't the model. Frontier models converge. What Anthropic got right is the connector architecture.
Cowork launched with MCP connectors for Google Drive, Gmail, Calendar, Slack, Notion, Linear, HubSpot, DocuSign, and a dozen more. Plugin marketplace. Private marketplaces for enterprises. Pre-built agents for HR, finance, engineering, ops.
The Auth0 pattern applied to enterprise AI. Don't build a better model. Build a better integration layer.
I saw this up close with the Anthropic sales team. Enterprise adoption isn't about demos. It's about fitting into existing workflows without asking anyone to change how they work.
The community play matters too. Claude Community Ambassadors program launched in March. User-organized meetups and hackathons worldwide. Funded by Anthropic, run by users. Hard to manufacture. Harder to compete with once it's running.
My prediction
By end of 2027, the winning enterprise AI platform won't be the one with the best model. It'll be the one with the most connectors to the tools companies already use.
Anthropic moved first and moved fast. Integrations create lock-in. Retraining costs compound. But "killed the game" implies the game is over. It's not. Codex has the developer trust. Gemini has the distribution. Meta has the open-source model layer. Mistral has sovereignty and global ambition. Microsoft has the distribution and now Anthropic's models.
The board just got set.