Anthropic published their latest growth numbers this week and the headline stat is wild: Claude’s paying consumer base has grown faster in Q1 2026 than in all of 2025 combined. Dario Amodei did a brief interview where he mentioned that Claude Pro subscriptions tripled since January, largely driven by word-of-mouth from developers who started recommending it to non-technical friends and colleagues. The product-market fit story seems real.

But I watched someone at a coworking space last Thursday use Claude for what should have been a simple task, and it looked like an elaborate ritual. She had claude.ai open in one tab, a Google Sheet in another, and a client portal in a third. The workflow was: read something in the portal, copy it, paste it into Claude, wait for Claude to process it, copy Claude’s output, paste it into the Sheet, go back to the portal, repeat. Alt-tab, Ctrl-C, Ctrl-V, over and over, for what turned out to be about 40 minutes of invoice data entry that the AI understood perfectly on the first try every single time. The bottleneck was never the model. It was the copy-paste shuttle between three browser tabs.

Millions of people doing the same alt-tab dance

Anthropic’s growth numbers confirm something most power users already suspected: Claude is genuinely good enough that people pay for it and keep paying. The retention curve Amodei shared shows paying users stick around at rates comparable to Netflix, which for a productivity tool is sort of absurd. But retention measures whether people come back to claude.ai, and that is a different question from whether people are getting full value from the model they are paying for.

The gap is obvious once you look for it. Claude cannot see your browser. It cannot read the page you are working on, cannot fill in the form you need completed, cannot extract the table sitting right there in your active tab. So every interaction with Claude that involves real browser work requires manual context transfer, which is a polite way of saying you become the integration layer between your AI and your actual work, copying text back and forth like a human API adapter.

Your tab is already authenticated

And there is a second problem hiding inside the first one. Most of the tasks people use Claude for involve websites where they are already logged in. Drafting a reply in Gmail, pulling data from a dashboard, navigating a CRM. But Claude on claude.ai has zero access to any of that session state, which means the user has to manually extract context from authenticated pages and re-describe it in the chat window, losing formatting, losing surrounding context, losing the ability for the AI to interact with the actual interface.

A browser agent flips this entirely. Dassi sits in Chrome’s side panel and sees the same page you see, running Claude (or GPT, or Gemini) against your actual browser context without needing OAuth tokens or API integrations for each service. Because it operates in your local browser with your existing sessions, there is no credential sharing, no third-party access grants, and no copy-paste. The model that Anthropic spent billions training finally gets to work with the information that was sitting right there the whole time.

The BYOK angle nobody talks about

What makes this especially strange is that Anthropic charges for Claude Pro, people happily pay, and then those same people manually shuttle context into a sandboxed chat window when the model could be running directly against their browser through a BYOK setup. You already own the key. The model is the same model, same weights, same capabilities. The only thing that changes is where it runs and what it can see.

I keep thinking about this in terms of the smartphone analogy, because in the early days of mobile apps a surprising number of people would use their phone’s browser to navigate to a website, copy a piece of information, open a different app, and paste it in. Then apps started talking to each other through share sheets and deep links and the copy-paste workflow vanished overnight. Not because the apps got smarter, but because the plumbing connected them. Browser agents are the same transition happening for AI tools, connecting the model to the surface where work actually happens instead of keeping them in separate boxes.

What Anthropic is not building

Anthropic has Claude artifacts, Claude Projects, the new Claude integrations with various developer tools. All excellent products in their own lanes. But nothing in their current roadmap (at least nothing public) addresses the fundamental problem that Claude cannot see your browser tabs or interact with the web pages you are working on. Their computer use API takes screenshots of an entire desktop, which is a blunt instrument that creates more privacy concerns than it solves and requires running a virtual machine for anything beyond demos.

So there is this odd situation where Anthropic builds the best reasoning model on the market, millions of consumers pay for it, and the actual deployment surface remains a chat window that is architecturally isolated from the user’s real workflow. The growth numbers are impressive and genuine, but they also represent a ceiling — every one of those paying users hits the same wall the moment they need Claude to do something with a web page rather than just talk about it.

Browser agents that run locally and let you bring your own API key are the obvious next step, and I suspect Anthropic knows this even if they are not the ones building it. The model is ready. The users are ready. The missing piece was always the plumbing between the chat window and the tab where the work lives, and that piece exists now. It’s just not where most people are looking.