Anthropic just killed OpenClaw. Your Claude Max subscription no longer covers third-party tools. Boris Cherny dropped the news on X on a Friday night (classic), and the community lost it. Everyone is screaming about the price.
They're screaming about the wrong thing.
I know because this is the third time it happened to me. January: a wall of 403s, my OpenClaw instance dead, me binge-reading Cherny's tweets like a maniac. I yelled at the screen. Everyone was yelling. February: I rebuilt. Kimi K2.5, fifteen dollars a month, pride of the guy who found the fire escape while the building burns. Today: I just nodded. No surprise, no anger. And that's when it clicked: the problem was never Anthropic. The problem is that we all believed an all-you-can-eat buffet for machines that eat 24/7 could last. The real trap isn't the price hike. It's what Anthropic wants you to switch to instead.
TL;DR: Flat-rate AI for agents was never a business model, it was VC subsidizing adoption. Anthropic is pushing you toward Cowork and first-party tooling, which is vendor lock-in dressed as optimization. But unlike telecom, AI models are interchangeable. The switching cost is near zero. Run env | grep ANTHROPIC before you do anything else.
The Gym Membership Model
There's a French saying: better pray to God directly than bother with the saints. I never trusted wrappers. Not for long, anyway.
But OpenClaw wasn't just a wrapper. It was an arbitrage play. A $200/month Claude Max subscription routed through a third-party harness that removed all rate limits. Power users were doing somewhere between one and five thousand dollars of compute on that subscription. A gym membership where you move your entire company's fitness program into the squat rack and expect the gym owner to smile about it.
The mobile carriers figured this out twenty years ago. "Unlimited data" was always a bet that most people would scroll Twitter on the train, not tether their laptop and download entire seasons of whatever. The bet held because humans sleep, eat, commute. AI agents don't. They run twenty-four hours, seven days, consuming tokens at a pace that makes the carrier math implode.
This is not an Anthropic-specific problem. Cursor went through the same "important pricing updates." Back in August 2025, an article on Medium called "The AI Subscription Illusion" already described the prisoner's dilemma perfectly: if one provider switches to usage-based pricing, they lose users to whichever VC-funded competitor still maintains the flat-rate illusion. Everyone knows the model is broken, nobody wants to blink first.
Every company that offered flat-rate access to frontier models eventually hits the same wall: the better the model gets, the more compute each agent session burns. The capacity cost scales with capability. The math doesn't care who you blame.
Now, to be fair: the flat-rate era wasn't useless. It let thousands of builders discover and adopt AI agents in the first place. Anthropic wasn't wrong to offer it. They were wrong to pretend it would last.
Nobody ships a buffet where the food gets more expensive to make every quarter.
The Trap Nobody's Discussing
Everyone on X is debating the price. Nobody is talking about the destination.
Anthropic is pushing you toward Cowork and Claude Code first-party. The official justification is technical: first-party tools optimize the "prompt cache hit rate," meaning they reuse previously processed text to save compute. Third-party harnesses like OpenClaw bypass those optimizations. Translation: you pay less, but you're locked into their compute pipeline, optimized for THEM, not for you.
I've been building with CLIs and open tooling long enough to know where this ends. Proprietary ecosystems always start with "we optimized it for you" and end with "you can't leave." The telecom playbook. The enterprise SaaS playbook. The cloud vendor playbook. Every single time.
And the first-party tooling has its own problems. Cowork shipped with prompt injection vulnerabilities that let attackers steal files via whitelisted curl commands within forty-eight hours of launch. That's the tool they want you to migrate to. The "optimized" alternative.
Wrappers aren't better, by the way. Lovable ran fifty-seven incidents in ninety days. Security researchers found sixteen vulnerabilities in a single hosted app. Users get charged for the AI's own debugging loops. Whether you go through the saints or through the church, the intermediary adds a failure mode you didn't ask for.
Anthropic has patched some of those Cowork issues since January. The point isn't that Cowork is dangerous today. The point is structural: when you depend on a single vendor for your model, your tooling, AND your compute pipeline, every bug is a single point of failure. Every pricing change hits you with no escape route. Every "important update" email is a hostage negotiation where you already gave up your leverage.
Your stack is only as independent as the vendor you can't replace.
What I Actually Pay (And One Command to Run Before You Switch)
Three things, quick.
First: Anthropic isn't slamming the door completely. They're installing a toll booth. Existing subscribers get a one-time credit equal to their monthly plan cost. There are discounted "extra usage" bundles if you want to keep your Claude login. And there's a refund option by email. You can keep going with Claude through extra usage or switch to an API key. Not a wall. A price increase with a grace period.
Second: what I actually pay. I already solved this problem in February, so I won't re-do the whole tutorial. I rebuilt the entire setup for $15/month with Kimi K2.5 as primary and MiniMax as fallback. Two $5 VPS instances for redundancy, works fine, ships every day.
Third, and this is the one nobody covers: the silent billing gotcha.
If you have ANTHROPIC_API_KEY set anywhere in your shell environment, claude -p will use it silently instead of your OAuth subscription. No warning, no prompt, no confirmation dialog. It just bills your API account. You think you're on your Max subscription, burning through your flat-rate allowance. You're not. You're on pay-per-token and the meter is running.
Run this right now:
env | grep ANTHROPIC
Or if you want to be thorough:
grep -r ANTHROPIC_API_KEY ~ 2>/dev/null
If either returns something and you thought you were on your subscription, you have a problem. Fix it before you do anything else.
(And yes, the service_tier and total_cost_usd fields in Claude's JSON output are misleading. They show data even on OAuth/Max where they don't reflect actual billing. That's a metadata/reporting issue, not a real charge. But a stray API key in your environment? That one's real.)
A ghost env var is all it takes to turn your "free" session into a bill.
Why AI Isn't Telecom (And Why You Win This Time)
The gym membership analogy breaks down at one specific point, and it breaks in your favor.
In mobile, the carriers won the war. Switching networks meant different coverage maps, slow number portability, early termination fees. The switching cost was enormous. You could complain about AT&T all you wanted. You were staying with AT&T.
In AI, models are interchangeable. I proved it three times: Claude to Kimi K2.5, same stack, same quality for agent workloads. OpenClaw already supports Codex, MiniMax, Z.AI, Gemma 4. The community is already voting with its feet. Google just released Gemma 4 under Apache 2.0 with day-one OpenClaw compatibility, and the dev community received it like a jailbreak 🔓. Running Gemma 4 locally on an RTX GPU means zero API costs. Zero vendor dependency. The model runs on your hardware.
The switching cost in AI is close to zero. That one fact changes everything about the power dynamic. Anthropic's vendor lock-in strategy cannot work long-term because the moment they make the cage uncomfortable, you walk into the next one. Or better: you stop using cages entirely.
The multi-model stack is the exit. The only question worth asking isn't "how much does it cost" but "how many providers can you lose before your pipeline stops."
In six months, Anthropic will have a new pricing tier with a corporate name and a blog post explaining why it's better for everyone. OpenAI will have copied it. "RIP flat-rate" with ⚰️ emojis.
Meanwhile, the people who build will have already moved. Not because they're smarter. Because they understood one simple thing: in a world where models are interchangeable, the only real risk is depending on a single provider.
How many providers do you have left if this one closes tomorrow?
Sources
Boris Cherny's X thread on the OpenClaw subscription cutoff (April 3, 2026). Amit, "The AI Subscription Illusion: Why Flat-Rate Pricing Is Failing" (Medium, August 2025).
(*) The cover is AI-generated. The lobbying budget for that gymnasium metaphor was surprisingly reasonable.
Top comments (0)