The language is new. The implications are not subtle.
Anthropic quietly updated its Claude Code legal and compliance docs to include this paragraph:
"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."
The policy explicitly names the Agent SDK — Anthropic's own tool for building AI agents — as off-limits for consumer plan OAuth tokens. If Anthropic's own SDK isn't exempt, third-party tools certainly aren't. The targets are obvious: OpenClaw, Pi Agent, and every other agent framework that lets users authenticate with their existing Claude subscription instead of paying separate API costs.
What Changed
Previously, the boundary between "Claude Code usage" and "third-party tool usage" was ambiguous enough that many OpenClaw users authenticated via OAuth tokens from their Max subscriptions. The Max plan at $200/month offered what felt like unlimited Claude access — why wouldn't you pipe that into your personal agent?
The updated docs draw a hard line: OAuth authentication is "intended exclusively for Claude Code and Claude.ai." Everything else requires API keys through the Claude Console or a supported cloud provider. Anthropic reserves the right to enforce "without prior notice," which means accounts could be banned before users even know the rule exists.
The Developer Reaction
The response on X has been swift and largely hostile.
"Anthropic giving bad vibes lately, I thought they were the cool guys?" wrote developer Flavio Copes. The sentiment was widely shared.
Ian Nuttall pointed to the competitive opening: "The Codex team need to get a Sonnet 4.6 speed and personality model shipped ASAP! Huge amount of Claude to Codex switches if they do."
Gergely Orosz, author of The Pragmatic Engineer, was blunt: "API cost for Anthropic is far too high vs other players, and so using the API just makes 0 sense to build on top of Claude. Seems Anthropic is happy to state they are good with having pretty much no ecosystem around Claude."
Developer Rhys Sullivan has already moved: "I've completely swapped over to 5-3-codex at this point for most things. The $20 and $200 ChatGPT plans are both really good."
The Economics Behind the Ban
The math explains the policy. A Max subscription costs $200/month for heavy Claude usage. API pricing for Claude Opus 4.6 runs $15 per million input tokens and $75 per million output tokens. An active OpenClaw agent running Opus can burn through millions of tokens per day — the kind of usage that makes a $200 flat-rate subscription deeply unprofitable for Anthropic.
One X user put it plainly: "It's pretty clear Claude Max subscription is economically unfeasible. The only reason it exists is to promote the Anthropic ecosystem to new users. Thus, people who use it for multiple accounts or for non-Anthropic products (e.g. OpenClaw) get cut off. It's just business."
It is just business. But it's business that pushes the most engaged users — the ones building agents, the ones evangelizing Claude's capabilities, the ones creating the ecosystem — toward competitors who price differently.
The Multiple Accounts Crackdown
The OAuth ban arrived alongside reports of Anthropic cracking down on users with multiple Max accounts. Developer Dwayne posted: "They're now going after people with multiple paid Max accounts. You're paying full price, multiple times, and they're treating you like a criminal."
Ed Zitron reported hearing from Enterprise Max account holders that Anthropic is moving them to API billing. The pattern is clear: Anthropic is closing every loophole that lets heavy users access Claude at flat-rate pricing.
What This Means for OpenClaw Users
If you're running OpenClaw with a Claude OAuth token from a Free, Pro, or Max plan, you have two options:
Switch to API keys. Create an account on the Claude Console, add a payment method, and use API key authentication. You'll pay per token, which for Opus 4.6 is significantly more expensive than a Max subscription for heavy usage.
Switch models. OpenClaw supports every major model provider. OpenAI, Google, Mistral, local models via Ollama — the framework is model-agnostic by design. The latest OpenClaw release even added Sonnet 4.6 support with forward-compatibility fallback.
The irony is hard to miss. OpenClaw's VISION.md — which we covered earlier today — lists "Supporting all major model providers" as a core priority. That design decision, made months ago, now looks less like a nice-to-have and more like insurance against exactly this kind of platform risk.
The Bigger Picture
Anthropic built the model that many consider the best in the world for agentic work. Claude Opus and Sonnet are the default choices for serious OpenClaw deployments. And now Anthropic is telling the ecosystem that grew around that capability: pay API rates or leave.
This is a familiar pattern in platform economics. The platform subsidizes adoption, an ecosystem forms around the subsidy, then the platform reprices and the ecosystem scrambles. What's unusual here is the speed — Claude Max has existed for months, not years, and the ecosystem it enabled is barely weeks old.
Whether this accelerates the shift toward OpenAI's Codex, toward open-source models, or simply toward grudging API payments remains to be seen. But the message from Anthropic is unambiguous: the Claude you build with is not the Claude you subscribe to. They are different products, at different prices, with different rules.
The developers who assumed otherwise just found out.