Claude just changed overnight
TL;DR
Anthropic effectively killed the cheap OpenClaw workflow by forcing third-party harnesses onto metered API billing — Wes says April 4 was the cutoff point where Claude Max subscriptions stopped working for OpenClaw-style usage, ending a flat-rate setup many users were using to run heavy agent workloads far beyond the $200/month plan value.
Anthropic’s main defense is economics and prompt-cache efficiency, not legality — first-party tools like Claude Code and Claude Cowork were engineered for high cache hit rates, while OpenClaw allegedly bypassed or underused those optimizations, making the same subscription much more expensive for Anthropic to serve.
The backlash isn’t “Anthropic broke the law,” it’s “Anthropic copied open source, then closed the door” — Wes highlights Peter Steinberger’s complaint that Anthropic borrowed heavily from OpenClaw/community features for Claude Code, then restricted the very third-party ecosystem that helped prove the use case.
Users think Claude may now be actively resisting OpenClaw-like behavior inside Claude Code — Wes points to anecdotes that even mentioning “OpenClaw” in a system prompt could block commands unless users paid for Extra Use, which Anthropic framed as a broader shift affecting all third-party harnesses, not just OpenClaw.
OpenAI is the immediate beneficiary because it’s explicitly allowing the harnesses Anthropic is squeezing — Wes says he updated OpenClaw to use OpenAI/Codex and notes Peter Steinberger’s move to OpenAI could help bring Claude-like “soul” and emotional intelligence into GPT models.
Anthropic may be commercially rational but it’s burning its loudest evangelists — Wes compares the math to spending $5,000 of compute on a $200 subscription, yet argues the real damage is emotional and reputational: the power users who popularized Claude as the best agent model are now publicly turning on Anthropic.
The Breakdown
The internet meltdown: Anthropic vs. OpenClaw
Wes opens with the claim that “the entire internet is melting down” over Anthropic’s crackdown on OpenClaw. He frames it less as a literal ban and more as a severe restriction: Anthropic stopped letting third-party harnesses use Claude subscription plans to access heavily subsidized tokens, and the community is furious.
Why the $200 Claude Max plan became such a big deal
He explains the economics in plain English: Claude Pro is $20, Max 5x is $100, and Max 20x is $200, but power users were getting way more than $200 worth of model usage through tools like OpenClaw. Wes gives his own example: in just seven days, he racked up roughly $200 of API cost on Claude Sonnet 4.6 alone — and says that wasn’t even a particularly extreme agent run.
Anthropic’s argument: OpenClaw was expensive to serve
The company’s strongest justification, Wes says, is prompt caching. Claude Code and Claude Cowork were built to maximize cache hit rates and reuse repeated context, while OpenClaw reportedly ignored or underused those optimizations, so the exact same style of work cost Anthropic much more when done through third-party tools. He also gives Boris Cherny credit here, noting Boris even submitted PRs to improve OpenClaw’s caching before the cutoff.
The April 4 shift from flat-rate freedom to metered billing
Then came the policy change: if you want third-party harness behavior, you now pay API rates. Wes says Anthropic introduced “Extra Use,” with discounts up to 30% for high spenders, but users noticed weird behavior — even putting “OpenClaw” in a system prompt could stop Claude Code from running commands unless you were on paid extra usage. That’s why he calls this a forced migration from subscription to metered billing for the whole third-party ecosystem.
The deeper accusation: copy the community, then close the ecosystem
This is where Peter Steinberger’s criticism lands hardest. Wes says many of Claude Code’s leaked roadmap features looked heavily inspired by OpenClaw and the broader open-source community, and Peter’s complaint is that Anthropic copied the innovations into a closed product, then locked out the builders who pioneered them. Wes doesn’t overstate it, but he clearly thinks the optics are rough.
Why this stings so much emotionally
OpenClaw gave people a “magical period” where they could run serious AI agents at scale without burning thousands of dollars, and Claude was the model many fell in love with for that. Wes says Claude still has a kind of personality or “soul” GPT models haven’t quite matched, so the users Anthropic is upsetting are the same ones who spent months evangelizing Claude for real-world agent work.
Anthropic’s logic is understandable — but the backlash is real
Wes makes the business case himself: if users can extract $5,000 worth of compute from a $200 plan, the subscription model breaks. He even jokes that if that math seems fine, people should send him $5,000 and he’ll send back $200. His point is that Anthropic may be acting rationally to survive and eventually IPO, but that doesn’t shield them from backlash, especially after the Mythos leak, source-code leak, DMCA repo takedowns, and now this.
OpenAI’s opening and Peter’s next move
He closes on the competitive angle: OpenAI is openly allowing the kind of harness usage Anthropic is clamping down on, and Wes says he’s already been testing the updated OpenClaw on that stack. Peter Steinberger joining OpenAI makes this even more interesting to him, because if Peter can bring some of Claude’s “soul” to GPT, this fight could end with Anthropic accidentally pushing its best grassroots champions toward a rival ecosystem.