Why did Anthropic ban OpenClaw? Here's the full story.

Building in Public · By Caleb Sakala · April 4, 2026

My cofounder messaged me this morning. "Claude just banned OpenClaw. Can they ban other products too? Can they ban us?"

He's not technical. He builds the business side. And his reaction is probably the same one most people had when they saw the headlines. Anthropic banned OpenClaw. Full stop. That's the version spreading on X right now.

It's also not what happened. And since most of the coverage either skips the technical details or buries them under outrage, I want to walk through the full picture. What OpenClaw was doing, how it worked, why Anthropic shut it down, and whether they were right to do it.

Short version: Anthropic closed a loophole that let people run industrial-scale AI agent workloads through a $20/month consumer subscription that was never built for that kind of usage.

If you want to understand what Anthropic actually did, forget about tech for a second and think about immigration policy.

Whatever you think of Donald Trump's politics, his claimed position on immigration is easy to summarize: he's not against people coming into the country. He's against people coming in through the wrong entrance. Legal immigration, visa process, front door: fine. Crossing the border without documentation, skipping the process, avoiding the system: not fine. The argument isn't "nobody gets in." The argument is "use the legal process."

That is exactly what Anthropic did here. OpenClaw users weren't banned from the building. They were told to stop climbing through the window and use the front door instead. The front door is the API. It works. It's open. You just have to pay the posted rate when you walk through it.

What everyone thinks happened

The story on X: Anthropic killed OpenClaw because they feel threatened by it. They want everyone on Claude Code and Claude Cowork, their own tools, so they're using platform power to crush an open-source alternative. Peter Steinberger, OpenClaw's creator, left for OpenAI in February. Weeks later, Anthropic goes after his project. Writes itself, right?

Clean story. Wrong story.

If Anthropic wanted to ban OpenClaw, they'd block API key access. They haven't. You can open a terminal right now, plug a Claude API key into OpenClaw, and run it. Nobody stops you. What Anthropic stopped is something very specific, and understanding it requires knowing how Claude access works under the hood.

Two doors into the same building

Anthropic sells Claude access through two separate systems.

The first is the API. You get a key from the Anthropic Console, you pay per token. Every word you send to Claude costs money. Every word Claude sends back costs money. Claude Sonnet runs $3 per million input tokens, $15 per million output tokens. Opus costs more. Your bill tracks exactly what you use. This is the system Anthropic built for developers and businesses who want to integrate Claude into their own products.

The second is consumer subscriptions. $20/month for Pro, $100 to $200/month for Max. These give you access through Anthropic's own apps: Claude.ai and Claude Code. Flat monthly fee, generous usage allowance.

These two systems have completely different economics. That difference explains everything about the OpenClaw situation.

How subscriptions survive at all

Anthropic is almost certainly losing money on heavy subscription users. Or breaking even at best.

Running large language models costs real money per request. Every token generated burns GPU compute. When Anthropic charges $3 per million input tokens through the API, that price reflects something close to actual serving costs, plus a margin.

A $20/month Pro subscription gives you way more effective usage than $20 of API credits would buy. How?

Part of it is engineering. Anthropic's own apps are built to be computationally efficient. Claude Code maximizes prompt cache hit rates. It reuses previously processed instructions so the same system prompts don't have to be recomputed on every request. That cuts the real compute cost per interaction by a lot. Anthropic built those optimizations into their own tool, and Anthropic captures those savings.

The other part is just statistics. Subscriptions work like gym memberships. Most people who pay $20/month don't use $20 worth of compute. They open Claude.ai a few times a week, ask some questions, close the tab. The light users subsidize the heavy users. Averaged across millions of subscribers, the economics hold. Barely.

That balance is fragile. It depends on most subscribers using less than their share of compute. And it depends on the power users accessing Claude through apps that are optimized to keep per-request costs low.

What OpenClaw did to that balance

OpenClaw is a third-party AI agent framework. Open source, widely used, designed to run agents that handle email, web browsing, code execution, and general automation. Many of those agents run around the clock.

The question is how OpenClaw connects to Claude.

The clean path is the API. Get a key, pay per token, run your agents however you want. Anthropic actively supports this. Boris Cherny, Anthropic's Head of Claude Code, even submitted pull requests to improve OpenClaw's prompt cache hit rates for API users.

But that's not what most OpenClaw users were doing.

They were using OAuth tokens. When you log into Claude Code with your Pro subscription, Anthropic issues an OAuth token that tells their servers "this person is a paying subscriber, let them through." Every request from Claude Code carries that token.

OpenClaw figured out how to use those same tokens. Its requests showed up at Anthropic's servers looking like they came from Claude Code. Same authentication, same credentials. But the actual traffic was coming from a completely different application running completely different workloads, with none of the prompt caching that makes subscription pricing work.

Put it this way. Anthropic built an employee entrance to their building. Keycard access, optimized for people who come in at predictable times, go to predictable floors, use predictable resources. The building's systems are planned around those patterns.

OpenClaw cloned the keycard. Now thousands of people are streaming through the employee entrance running 24/7 operations out of the building, pulling resources at rates nobody planned for. They're not employees. They don't follow employee patterns. The infrastructure is straining under load it was never designed to handle.

Anthropic didn't ban those people from the building. They said: use the front door and pay the posted rate. The front door works fine.

The math that broke

The subsidization problem becomes concrete when you run the numbers.

An OpenClaw agent running a multi-step workflow might consume 50,000 tokens per execution. An agent that runs throughout the day could fire off hundreds of those. At API rates on Sonnet ($3/$15 per million tokens), that costs real money. Hundreds, sometimes thousands of dollars per month depending on volume.

Through a $20 Pro subscription, all of that was effectively included. And because OpenClaw doesn't trigger the same prompt caching that Claude Code does, Anthropic was absorbing the full, unoptimized compute cost of every single token. Not the reduced cost their subscription math assumes.

Scale that across thousands of users. The gym membership model falls apart when everyone shows up daily and runs on the treadmill for eight hours. Except these treadmills cost Anthropic real money for every minute they're in use.

Boris Cherny said it directly: "Our subscriptions weren't built for the usage patterns of these third-party tools." That is not corporate spin. That is arithmetic.

So was Anthropic right to do this?

I get why people are upset. They had a good thing going. The good thing is ending. That hurts if you're an independent developer who built workflows around $20/month covering your agent costs.

But that deal was never going to last.

The $20 price point was always subsidized. You were getting more compute than you paid for. That arithmetic works when usage is moderate and runs through software Anthropic optimized for efficiency. It does not work when third-party tools route continuous, heavy, unoptimized agent workloads through a channel designed for lighter, cached interactions.

Anthropic is not saying "stop using OpenClaw." They are saying "if you're going to run agents generating AI output around the clock, pay for the compute you're consuming." That is how every utility works. Your electricity provider gives you a flat rate for normal household usage. Start running a bitcoin mine out of your garage and they will notice, and they will charge you differently.

The people calling this a walled garden play are mixing up two different things. A walled garden means you can't leave. Nobody is stopping you from leaving. OpenClaw works with OpenAI, Google, DeepSeek, Qwen, and dozens of other providers. You can switch models in a config file. What Anthropic is saying is narrower: if you use our models, pay the rate that reflects what it costs us to serve you. That is a pricing correction, not a competitive play.

Every AI subscription has this same problem

This is not an Anthropic-only issue.

Every major AI provider offers consumer subscriptions that are subsidized to some degree. OpenAI's $20/month ChatGPT Plus, Google's Gemini Advanced. These prices exist to drive adoption. They do not reflect the real cost of heavy usage.

The second any third-party tool makes it easy for power users to consume 10x or 100x what a typical subscriber uses, at the same flat rate, the provider faces exactly this problem. OpenAI publicly endorsed using Codex subscriptions with third-party harnesses the same day Anthropic announced the crackdown. Smart competitive move. Also a bet that the usage patterns won't torch their margins the way they did Anthropic's. If OpenClaw's user base migrates to Codex subscriptions with the same volume, don't be surprised when OpenAI makes the same call six months from now.

Flat-rate AI pricing was always a customer acquisition play. Not a permanent business model.

What to do about it

If you're running OpenClaw or any third-party tool with Claude, here's the situation.

Get an API key from the Anthropic Console. Pay per token. You'll know exactly what your costs are. No surprises, no Friday-night emails changing the terms overnight.

Or stay on your subscription and use Anthropic's Extra Usage bundles. They're offering up to 30% off pre-purchased usage, plus a one-time credit equal to your monthly subscription cost, redeemable by April 17. Not nothing.

Or switch providers. OpenClaw works with nearly every major model through OpenAI-compatible endpoints. The model is a commodity input. The tools that survive long term are the ones that treat it that way.

For anyone building automation that matters, the real takeaway goes past this one incident. Running critical workflows on a consumer subscription, through an authentication path the provider says you shouldn't use, was never a stable foundation. It was an arbitrage with an expiration date. The date came.

Build so the model is a swappable component. Your workflow logic, your integrations, your tool connections: those are the parts that last. Each step in a Chase Agents workflow specifies its own model and provider independently. When Anthropic changed terms, workflows using Claude through API keys kept running. Nothing broke because the architecture was built for exactly this kind of provider-level shift.

The providers will keep tightening. The teams that build around that reality today will still be running when the next one makes the same move.