I’ve been a happy Anthropic Max subscriber for a while now. $200/month for their top-tier plan, which gives me access to Opus 4.6 — one of the best models available today. I use it to power Demerzel, my OpenClaw agent. In case you haven’t heard of OpenClaw, its an open-source AI assistant platform that went viral earlier this year. OpenClaw acts as a “harness” around AI models, bringing persistent memory, tool use, messaging integrations, and autonomous agent capabilities to users. OpenClaw includes a “provider” for wiring up your OpenClaw agents to Claude by piggybacking on your flat-rate subscription.
Well, the party is over, folks. Anthropic announced that third-party harnesses like OpenClaw will no longer be allowed to leverage Claude subscriptions. Instead, they’ll have to pay by the token. Ouch. First-party apps from Anthropic like Claude Desktop, Claude for mobile, and Claude Code are, of course, exempted from this policy, and will continue to provide generous usage limits.
The community reaction was swift and vocal. Angry outbursts on Reddit. Chaotic social media threads. Flooded Discord servers. I get it! We were all having so much fun! Boundless token consumption, baby! Let’s go!
But….
Anthropic was Right
If you’re a Claude user, you’ve probably noticed that Anthropic has had trouble keeping ahead of surging demand, and one of the biggest culprits driving that expansion is OpenClaw. Opus 4.6 is fantastically good, but its also resource intensive to run. Allowing users of a third-party open source project unfettered access created disruption for their mainstream users, and critically, their entire business model.
Anthropic is in the business of selling subscription products to end users at a fixed rate, and API access to developers metered by tokens. OpenClaw was exploiting a loophole that effectively allowed them to bypass the consumption-based API services by masquerading as an end user.
Let me explain with my own numbers. I tracked my actual usage through OpenClaw for a typical month. Between continuous heartbeat monitoring, proactive email and calendar checks, Claude Code supervision sessions, research tasks, and the general always-on nature of an autonomous assistant, my consumption-based cost would come out to over $2,000 per month — on top of the $200 base subscription.
That’s not a rounding error. That’s an order of magnitude more than the subscription itself.
And this is exactly the problem. Flat-rate subscriptions for frontier AI models only work when usage stays bounded — a human typing prompts in a chat window, a software developer working inside of Claude Code. These subscriptions allow their customers to leverage Anthropic’s models in predictable ways with reasonable limits.
But when you wrap a model in an always-on harness that’s making thousands of API calls around the clock, delegating to sub-agents, and autonomously chaining tool calls together, the underlying compute cost explodes.
Anthropic was essentially subsidizing power users like me to the tune of thousands of dollars per month. That’s not sustainable for any business, and pretending otherwise would have been worse for everyone in the long run.
Open Source Strikes Again
Open source software is now a mainstream phenomenon, but that wasn’t always the case. Back in the late 1990’s (I’m so old…), I was a teenager building software in the enterprise healthcare market. The hospitals we worked with had their own infrastructure running high-end commercial UNIX operating systems like AIX, SCO Unix, Solaris, and HP-UX. These operating systems were incredibly expensive, only moderately compatible with one another, and often ran on overpriced hardware from the OS vendor.
Then, in the early 2000’s, Linux started to make waves. A free and open source UNIX-ish kernel had the potential to disrupt the commercial UNIX market, and RedHat was the first company to make that disruption a reality. Before long, Linux became the operating system of choice for enterprise applications, and supported distributions from RedHat, SuSE, and Canonical rapidly displaced commercial UNIX. Open source changed an entire market.
While it may be a bit hyperbolic for me to compare Anthropic’s smackdown of OpenClaw to the rise of Linux, I think its a valuable thought experiment. In just a few short months, a community developed open source project was able to completely disrupt a company with billions of dollars of funding. Open source is so metal!
What’s Next?
Anthropic didn’t make an enemy here. They made a savvy decision that keeps them laser focused on their core customers. I’m still a happy Claude Max subscriber, and a daily user of Claude Code.
For my OpenClaw instance, I have tested a variety of alternatives, and landed on z.ai. Their GLM-5.1 model is priced attractively for high-volume usage and compares favorably to Opus 4.6 in day-to-day work. The transition was surprisingly smooth — my assistant runs the same, just backed by a different family of models.
The one trade-off? GLM-5.1 has a 200K token context window versus Opus 4.6’s ability to burst up to 1M tokens. In practice, that means my assistant can be a little forgetful on longer sessions — which, as it turns out, led me down a fascinating rabbit hole…
