Saw the Hacker News thread this morning: “Anthropic says Claude Code subscribers will need to pay extra for OpenClaw usage.” Not a rumor, not speculation. An actual surcharge bolted onto a subscription people were already paying $20/month for. And the comments section was exactly what you’d expect — a mix of resignation and low-grade fury from developers who adopted Claude Code specifically because it felt like a complete package.

I do not think Anthropic is being greedy. Running cloud browser infrastructure is genuinely expensive, and OpenClaw’s compute requirements are substantial when you’re spinning up headless Chrome instances at scale. But the surcharge reveals something that’s been true about cloud browser agents for a while now, and that most people haven’t fully internalized yet: the costs compound in ways you do not see coming.

The stacking problem

Your Claude Code subscription covers the LLM. OpenClaw adds a surcharge for browser orchestration. But those are just the line items you can see on a bill. The hidden costs are worse.

Every cloud browser task involves spinning up a remote instance, screenshotting the page, sending that image to a vision model, waiting for the model to reason about what to click, executing the click, screenshotting again, and repeating this loop until the task completes. A simple form fill that takes you 30 seconds might run through eight or nine round-trips, each burning compute and API tokens on Anthropic’s infrastructure. When people complain about OpenClaw being slow, they’re experiencing the cost as latency, which is just cost denominated in time instead of dollars.

And then there’s the authentication overhead I keep writing about. Because cloud browser agents start from a blank session every time, they burn additional cycles just logging into services you’ve been logged into for weeks. That is compute you are paying for that produces zero value — it just gets you back to the starting line that your local browser never left.

Your browser is already running

Something that gets lost in all the pricing discourse: you already own a perfectly good browser. It’s open right now, probably with a dozen tabs, each one authenticated to a different service. Your Gmail session, your Jira board, your Notion workspace, your CRM. All of that session state exists for free, maintained by Chrome’s cookie jar, costing you nothing beyond the electricity to run your laptop.

A local browser agent plugs into that existing infrastructure. Dassi lives in Chrome’s side panel and reads the same DOM you’re looking at, with your logins intact, no remote instance required. The only cost is whatever LLM tokens you consume, and because you bring your own API key, you’re paying the model provider directly at their published rate. No platform margin. No surcharge. No surprise line items three months from now when usage tiers change.

Subscriptions breed surcharges

This is the pattern that SaaS established and AI tooling is now repeating with alarming speed. You subscribe to a platform, the platform becomes load-bearing in your workflow, and then the platform introduces usage-based pricing on top of the subscription because the unit economics don’t work at the flat rate they originally quoted. We watched Heroku do it, we watched Figma almost do it before they reversed course, and now we’re watching it happen with AI agents in real time.

The BYOK model sidesteps this entirely. When you own the API key, the pricing relationship is between you and your LLM provider — Anthropic, OpenAI, Google, whoever you choose. If Claude gets expensive, you switch to GPT-5.2 or Gemini 3.1 Pro. The agent layer, if it’s local, has no cloud costs to pass through and no reason to add surcharges because it is not running anyone’s infrastructure but yours.

What the surcharge actually tells you

Anthropic pricing OpenClaw separately is an honest acknowledgment that cloud browser orchestration is a different product with different economics than a coding assistant. I respect the transparency, even though the result is annoying. But it should also make you ask whether cloud browser orchestration is the right architecture for your use case in the first place, because you’re now paying for both the intelligence and the plumbing when the plumbing was supposed to be the easy part.

Your Chrome installation is free plumbing. It handles rendering, authentication, session persistence, and DOM access at speeds that no cloud relay can match. Building an agent that leverages what already exists locally is just a fundamentally cheaper architecture than rebuilding the entire browser stack in somebody else’s data center and then billing you for the privilege.

The OpenClaw surcharge is $X/month today. It will be $Y/month in six months when Anthropic discovers the real infrastructure cost curve. And you will pay it, because by then your workflow depends on it, which is exactly how lock-in works with AI tools.

Or you could just use the browser you already have open.