I saw the Hacker News thread about Anthropic’s new OpenClaw pricing hit the front page yesterday, and the comments were about what you’d expect — a mix of frustration and resignation. Claude Code subscribers who’ve been using OpenClaw for browser automation are now looking at an extra charge on top of what they already pay for the model itself.

Anthropic’s framing is reasonable enough: OpenClaw requires compute for orchestration, browser infrastructure, and tool execution that goes beyond normal API usage. Fair. Running headless browsers in the cloud costs real money. But the result for users is a familiar one — you’re paying for the model, and now you’re paying again for the scaffolding around it.

The plumbing has a price tag

This pattern keeps repeating across the AI tools landscape, and it’s worth stepping back to notice how consistent it is. You subscribe to an LLM. Then you subscribe to a tool that calls that LLM on your behalf. The tool adds orchestration, maybe some browser infra, maybe a managed execution environment, and charges you a monthly fee for that layer.

OpenClaw’s new pricing is just the latest instance, but you can see the same structure in Operator, in various MCP hosting platforms, in managed agent services from startups that raised their Series A last quarter. The model is a commodity (increasingly), but the plumbing around it — the thing that actually connects the model to your work — that’s where the margin lives now.

And I get why companies build it this way. Cloud infrastructure is expensive, browser pools need maintenance, and somebody has to pay for the servers that keep your headless Chrome instances warm at 3 AM. So the cost gets passed to users as a second subscription layered on top of the first.

What you’re actually paying for twice

The thing that bugs me about this is how much of that infrastructure exists to solve a problem that doesn’t need solving if you rethink the architecture. OpenClaw spins up cloud browsers because it needs a browser to work in. Your computer already has one. It’s open right now, probably with a dozen tabs you’re actively using, logged into every service you care about.

Cloud browser agents — OpenClaw included — cannot see those tabs. They run in isolated cloud environments, which means they need fresh login credentials, session tokens, or API integrations for every service they touch. That isolation is what creates the infrastructure cost, which becomes the agent tax on your bill.

A browser agent that runs locally in your actual Chrome skips all of that. No cloud browser pool. No session management layer. No orchestration service sitting between you and the model. Your browser is the execution environment, and it’s already running, already authenticated, already showing the page you need help with.

BYOK kills the second subscription

Dassi works this way — it’s a Chrome extension that lives in your browser’s side panel, sees what you see, and acts on the page you’re looking at. You connect your own LLM (Claude API key, OpenAI key, or even just your existing ChatGPT subscription with no API key at all) and that is the entire cost structure. There’s no agent tax because there’s no cloud plumbing to fund.

I keep coming back to a question that seems obvious but apparently isn’t: if you’re already paying $20/month for ChatGPT or Claude, why would you pay another $X/month for a service whose main job is calling that same model on your behalf? The model does the thinking. The browser does the acting. What exactly is the middle layer contributing that justifies its own line item on your credit card statement?

For some use cases, the answer is genuinely “a lot” — if you need agents running unattended overnight, or browser sessions in geographic regions you’re not in, cloud infra earns its keep. But for the vast majority of daily browser work (drafting emails, filling forms, pulling data from dashboards, navigating CRM interfaces), the agent tax is paying for infrastructure you do not need because your own browser already provides it.

The pricing trend is only going one direction

Anthropic is not going to be the last company to add usage-based pricing to agent features. Every cloud AI platform that offers tool use, browser control, or autonomous task execution is going to find that the compute costs don’t fit neatly into a flat subscription, and the overage charges and premium tiers will follow. We saw SaaS companies discover this exact same monetization playbook over the past decade — give away the base product, charge for the integrations and automations that make it actually useful.

The difference with AI agents is that the “integration layer” is often just a model calling a browser, which is something your laptop can do for free right now while you read this sentence.

So when you see the next pricing announcement from whatever agent platform raises their rates next month, the question worth asking is not “is it worth the price” but “is the architecture that requires this price the only option.” Because increasingly, it isn’t. Browser agents that bring your own key and run in your own Chrome are a damn sight cheaper, and for most daily workflows, they do the same job.

The model is the expensive part. Everything else is plumbing. And you probably already own the pipes.