Anthropic and Google are banning OpenClaw users — here's why

By Linas Valiukas · March 27, 2026

In January, Anthropic quietly deployed a server-side block that killed OpenClaw's ability to use Claude subscription tokens. No announcement. Users just woke up to 403 errors and a vague terms-of-service notice. A month later, Google went further — disabling paying subscribers' Antigravity accounts without warning, mid-billing cycle, no refunds offered.

If you self-host OpenClaw and route it through a subscription API, this is a risk you need to understand. Two of the three biggest LLM providers have now taken action. The third hired OpenClaw's creator.

What Anthropic did

On January 9, 2026, Anthropic blocked subscription OAuth tokens from working outside their official Claude Code CLI. If you had Claude Pro ($20/month) or Claude Max ($200/month) and were routing requests through OpenClaw, your setup broke. The error message: "This credential is only authorized for use with Claude Code and cannot be used for other API requests."

There was no blog post. No email. Just a server-side change and silence for five weeks.

On February 17, Anthropic finally published documentation making it official. The key line: "Using OAuth tokens obtained through Claude Free, Pro, or Max accounts in any other product, tool, or service — including the Agent SDK — is not permitted and constitutes a violation of the Consumer Terms of Service."

Thariq Shihipar, an Anthropic engineer, told The Register that third-party harnesses "create problems for users and are prohibited by our Terms of Service." They also said they wouldn't be canceling accounts — the block was on the token type, not the user.

Small comfort if your entire automation stack just stopped working.

The economics behind it

Anthropic's reasoning is straightforward once you see the numbers. Claude Opus costs $15 per million input tokens and $75 per million output tokens at API rates. A $200/month Max subscription gives you access to the same model. An OpenClaw agent running autonomously — checking emails, responding to messages, managing calendars — burns through tokens fast. A single "how are you?" through OpenClaw consumes roughly 30,000 tokens when you include the system prompt and context.

Run that agent for a full day and you're looking at $1,000-$5,000 in API-equivalent costs on a $200 subscription. Some users were burning millions of tokens in a single afternoon. Anthropic was losing money on every one of them.

Their ToS Section 3.7 already prohibited accessing services "through automated or non-human means, whether through a bot, script, or otherwise" except via API key. They were technically within their rights from day one. They just waited until the financial bleeding got bad enough.

What Google did (and it was worse)

Anthropic blocked tokens. Google blocked people.

Starting around February 11, 2026, Google's automated systems began flagging OpenClaw OAuth traffic as "malicious usage" and disabling Antigravity accounts. No warning email. No grace period. Subscribers paying $250/month for AI Ultra — some on annual plans — logged in to find their access gone and a one-line message about a ToS violation.

It wasn't a handful of people. At least 25 users reported bans in a single Google AI Developer Forum thread. One company had half their Pro accounts disabled simultaneously. GitHub issue #14203 filled up with "yeah i got banned" replies.

Varun Mohan from Google DeepMind explained to The Register: "We've been seeing a massive increase in malicious usage of the Antigravity backend that has tremendously degraded the quality of service for our users." He said they had "limited capacity and want to be fair to our actual users."

Calling paying subscribers "not actual users" is a choice. A $250/month subscriber running OpenClaw and a $250/month subscriber typing into Gemini manually are paying the same amount. The difference is how many tokens they consume. Google's flat-rate pricing was a bet that most people wouldn't max it out. OpenClaw users did.

The fallout

Google took about twelve days to acknowledge the problem publicly. Another four to start restoring accounts. During that time, banned users had no self-service appeal, no clear path to resolution, and in some cases were still being billed.

A Google exec later clarified that only Antigravity access was blocked — Gmail, Drive, and other services weren't affected. That calmed some nerves, but the initial confusion was real. When Google disables something on your account, the first thought is "did I lose everything?"

Google eventually updated their FAQ to explicitly name OpenClaw, Claude Code, and OpenCode as prohibited third-party tools. They recommended switching to a Vertex or AI Studio API key for agent use — pay-per-token, not flat-rate.

The OpenClaw maintainers' response was blunt. Vincentkoc commented on the GitHub issue: "Some providers have TOS which may be violated when using your agents with those providers." And: "Sucks some peoples accounts have locked." The issue was closed as "won't fix."

Who isn't banning anyone (and why)

OpenAI. They hired Peter Steinberger — OpenClaw's creator — on February 14. Thibault Sottiaux from OpenAI publicly endorsed third-party harness use with their subscriptions. While Anthropic and Google were shutting doors, OpenAI was using OpenClaw compatibility as a selling point for ChatGPT Pro.

DeepSeek hasn't taken action either. Neither have Mistral or Groq. The bans created a migration wave — one user on Medium rebuilt their entire Anthropic-based setup on Kimi K2.5 and MiniMax M2.5 for $15/month. Others switched to local models through Ollama.

George Hotz called Anthropic's move a huge mistake, arguing it "will not convert people back to Claude Code, you will convert people to other model providers." DHH called it "very customer hostile." Steinberger himself called Google's bans "pretty draconian" and floated dropping Antigravity support entirely.

What this means if you self-host

The ban wave exposed a dependency most self-hosters hadn't thought about. You own the server. You control the config. But you don't own the API. And the API provider can cut you off.

If you're running OpenClaw with a pay-as-you-go API key, you're fine for now. Neither Anthropic nor Google has restricted standard API access. The bans targeted subscription OAuth tokens specifically — flat-rate plans being exploited for unlimited agent use.

But "fine for now" has limits. Three risks to think about:

• Runaway costs. A pay-as-you-go API key with an autonomous agent is an open credit line. If your agent decides to process a backlog or gets stuck in a retry loop, you'll find out when the invoice arrives. Claude Opus at $75 per million output tokens adds up fast when nobody's watching.
• Key exposure. Over 40,000 OpenClaw instances are running without authentication. If yours is one of them, your API key is sitting in a plaintext config file accessible to anyone who finds the instance. Someone else runs up your bill, and the provider flags your account for abuse.
• Policy shifts. Today the line is drawn at subscription tokens. Providers can move that line. If agentic usage starts costing more than expected on pay-as-you-go plans too, new restrictions aren't hard to imagine. Anthropic's ToS already prohibits "automated or non-human means" broadly.

The real cost isn't the API bill

What happened to the people who got banned wasn't just an inconvenience. Their automations stopped. Messages went unanswered. Scheduled tasks didn't run. If you'd built business processes on top of OpenClaw with a Claude Max subscription — email handling, customer responses, appointment management — all of that died the moment the token stopped working.

The Medium user who rebuilt on Kimi and MiniMax spent days migrating. The Google Ultra subscribers spent weeks in limbo waiting for account restoration. One user in the GitHub issue mentioned their OpenClaw instance held credentials for Telegram, Slack, and email — when Antigravity went down, they scrambled to check whether those credentials were still safe.

That's the part self-hosters underestimate. It's not about whether your current setup works today. It's about what happens when a provider changes their mind and you have to rebuild everything from scratch on a Saturday morning.

How managed hosting handles this differently

A managed OpenClaw service deals with API access on your behalf. The credentials are properly configured, ToS-compliant, and not your problem. If a provider changes their terms, the service migrates. If a model gets rate-limited, the service adjusts. You don't get a 403 error at 2 AM because a token type you didn't know existed got deprecated.

TryOpenClaw.ai manages API access, applies security patches same-day, and keeps your instance running with ToS-compliant credentials. Starts at $39/month — less than the cost of one month on Claude Max, and you won't wake up to a wall of 403s.

Frequently asked questions

Why is Anthropic banning OpenClaw users?

Anthropic blocked subscription OAuth tokens from working outside Claude Code. They didn't ban accounts — they killed the token type. A $200/month Claude Max subscription was being used to run $1,000-$5,000 worth of agentic workloads through OpenClaw. Their ToS already prohibited automated access via subscription tokens; they just started enforcing it.

Why is Google banning OpenClaw users?

Google's automated systems flagged OpenClaw traffic as malicious and disabled Antigravity accounts without warning. At least 25 users reported bans in a single forum thread, including $250/month AI Ultra annual subscribers. Google later restored accounts and updated their FAQ to explicitly name OpenClaw as a prohibited tool.

Can I still use OpenClaw with a pay-as-you-go API key?

Yes. Both Anthropic and Google's restrictions target subscription OAuth tokens, not standard API keys. If you use a pay-as-you-go key and pay per token, you're not violating either provider's terms. Google's updated FAQ recommends Vertex or AI Studio API keys for third-party agent use.

Did OpenAI ban OpenClaw users too?

No. OpenAI hired OpenClaw's creator and publicly endorsed third-party harness use with their subscriptions. DeepSeek, Mistral, and Groq haven't taken action either. The bans are specific to Anthropic (subscription tokens) and Google (Antigravity accounts).

What happens if my API key gets revoked while OpenClaw is running?

Your OpenClaw instance keeps running but can't make LLM calls. Every automation that depends on the AI model stops — messages go unanswered, scheduled tasks fail, and anything mid-conversation gets dropped. Your other connected services (Telegram, Slack, email) aren't directly affected, but the agent can't use them without a working LLM connection.