The Buffet Is Closed: Anthropic Just Locked Out Every AI Agent Power User
Anthropic sent the email Friday evening. Nobody reads emails on Friday evenings, which is probably the point.
Starting today at 3PM ET, Claude Pro ($20/month) and Max ($100-200/month) subscribers lost the ability to connect their plans to third-party AI agent tools. OpenClaw. Other harnesses. Anything that isn't Claude Code or Claude Cowork.
The official line from Boris Cherny, Head of Claude Code at Anthropic, posted on X: "We have been working hard to meet the increase in demand for Claude, and our subscriptions were not built for the usage patterns of these third-party tools. Capacity is a resource we manage thoughtfully and we are prioritizing our customers using our products and API."
That's the nice version. Here's the other version.
The Technical Cover Story
Anthropic's first-party tools, Claude Code and Claude Cowork, are built to maximize "prompt cache hit rates." This means they reuse previously processed text instead of recalculating it from scratch every time. It's genuinely clever and it saves a lot of compute.
Third-party harnesses like OpenClaw don't do this as efficiently. They tend to process things fresh each time. This means for every OpenClaw user burning through tokens, Anthropic is eating more compute than they would if that same user was running Claude Code directly.
Cherny confirmed this on X: "Third party services are not optimized in this way, so it's really hard for us to do sustainably."
He also noted he personally submitted PRs to improve OpenClaw's cache hit rate. Which is a nice gesture from someone who then helped kill the thing those PRs were meant to improve.
The compensation package: a one-time credit equal to your monthly plan price, good until April 17. And a 30% discount on new "extra usage" bundles if you want to keep going.
The Timing Questions
Peter Steinberger created OpenClaw. In February 2026, he joined OpenAI.
Last week, he posted: "Funny how timings match up. First they copy some popular features into their closed harness, then they lock out open source."
He's not wrong. Anthropic recently shipped Claude Code Channels, which lets you message agents through external services like Discord and Telegram. That feature existed in OpenClaw first. Now the version that existed in OpenClaw first is in Claude Code, and OpenClaw itself is getting squeezed out.
Steinberger also said he and OpenClaw board member Dave Morin attempted to talk Anthropic out of this. Their victory: delaying enforcement by one week. The policy was originally planned for March 28. They got it pushed to April 4.
That's not a win. That's a one-week heads up for the developer community.
The Real Cost
Here's the number that put things in perspective for me.
A single OpenClaw agent running autonomously for one day can burn $1,000 to $5,000 in API costs. That's not an exaggeration. When you're running an agent that can manage your email, your calendar, your flights, your reminders, all day long, the token usage adds up fast.
Anthropic was subsidizing that. Pro and Max plans have flat-rate limits that don't come close to covering what a full-day agent run actually costs Anthropic in compute. Every power user running OpenClaw was essentially getting compute at a steep discount, because the subscription model wasn't designed for how agents actually consume tokens.
Now the discount is gone.
Small builders are getting hit hardest. The founder of Telaga Charity posted: "If I switch both [OpenClaw instances] to an API key, it's going to be far too expensive to make it worth using. I'll probably have to switch over to a different model at this point."
That's the real story. The enterprises will be fine. They'll pay the API rates or negotiate enterprise deals. The individual builders, the indie hackers, the small shops that built their workflows around OpenClaw and a Pro subscription — they're the ones getting priced out.
What This Actually Means
We're watching a company make a cold calculation about margins versus growth.
Anthropic is betting that controlling the harness layer is worth more than keeping the power-user community happy. They're prioritizing the Claude Code experience, collecting telemetry on how users actually interact with their models, and managing rate limits at the first-party level.
The argument is rational. The subscription model was never built for autonomous agents running 24/7. Those usage patterns don't fit cleanly into a $20/month flat rate. At some point, the math breaks.
But the timing, the OpenAI hiring, the Claude Code Channels feature launch right before the crackdown — it doesn't pass the smell test for me. Steinberger's frustration is visible and it's earned.
The Bottom Line
This is what the AI industry's "sustainable growth" phase looks like in practice. Not everyone gets to be the customer Anthropic is optimizing for.
If you were running OpenClaw on a Pro plan, you have some options: pay-per-token API, the discounted extra usage bundles, or switch to a different model. None of them are as clean as what you had yesterday.
The era of subsidized, flat-rate agent compute is over. Anthropic just closed the buffet.
The question now is which lab opens the next one.