If you’ve spent any time in WhatsApp groups or Discord servers where vibe coders hang out, you’ll have noticed something. People are furious.
Not mildly annoyed (genuinely, keyboard-smashingly furious) because Anthropic had the audacity to limit how many tokens they can burn through on a flat-rate subscription.
One developer on a $200-per-year Claude Pro plan reported that their account maxes out every Monday and resets Saturday… meaning they get roughly 12 usable days out of 30. Outrageous. Scandalous. The end of innovation as we know it. 😬
Or, you know, the completely predictable consequence of treating an AI model like an all-you-can-eat buffet and then being surprised when the restaurant starts asking why you’ve been there for 14 hours. 🤷♂️
Let me be clear about something before the pitchforks come out: Anthropic has been opaque about this, and that’s genuinely frustrating. The Pro plan promises only “at least five times the usage per session compared to the free service”… no hard token numbers, no published limits, just a vague multiplier.
That’s not good enough for anyone trying to plan a real workflow. Fair criticism, and worth making loudly.
But that’s a different conversation from the one I want to have here.
Because buried underneath all the outrage is a question nobody seems to want to ask: what exactly were people building with all those tokens?
Why Anthropic’s Move is Great for the Ecosystem
The All-You-Can-Eat Problem
With the rise of agentic AI functionality like vibe coding and “computer use” capabilities, flat-rate AI subscribers are burning far more tokens than ever before, and the big AI providers are struggling to keep up with demand. Anthropic eventually discovered something remarkable: some users were consuming tens of thousands of dollars in compute while paying just $200 monthly subscriptions. Not a handful of outliers. Enough of them to break the economics of the entire service.
Inside the vibe coding community, those users were celebrated. A leaderboard called Viberank emerged, transforming what was once private spending into public competition, with developers racing to the top based on token consumption.
Think about that for a second. A status symbol built entirely on how much compute you could torch. Not what you shipped. Not the revenue you generated. Not the problem you solved. Just the raw volume of AI resources you incinerated on your way to… somewhere. 🤦♂️
One Viberank top spender ran sessions costing $500 each and used eight shared accounts to relay session history into new contexts just to keep the work going. The project got stars on GitHub. Whether it generated any revenue, nobody mentioned.
The Filter Nobody Asked For (But Everyone Needed)
Here’s my contrarian take:
Anthropic raising limits and tightening access is the most useful thing that could have happened to this ecosystem right now.
Building with AI is cheap enough that the barrier to starting is basically zero.
That’s the point of democratised access, anyone can build, great. The problem is that zero friction at the start means zero thinking at the start. People spin up vibe coding sessions to clone tools that already exist, to prototype ideas they’ve never stress-tested against real demand, to build things they describe as “projects” but are really elaborate procrastination dressed up in a GitHub repo.
When tokens are effectively free, that behaviour is invisible. When tokens cost money, it becomes very visible very quickly. Suddenly, you have to ask: Is what I’m building worth paying for?
That’s not a question the vibe coding community wants to sit with, because the honest answer (for a significant chunk of what gets built) is NO. And that’s fine. Experimentation has value. But experimentation is not a business model, and free compute is not a signal of viability.
If your project can’t survive paying actual market rates for the compute it consumes, your project is not sustainable. That’s not a controversial position. That’s just arithmetic. 🤷♂️
The Environmental Maths You’re Ignoring
There’s another dimension to this that almost nobody in these WhatsApp groups is talking about. In 2025, AI-generated code emits up to 19 times more CO2 than human-written code during development, and that’s before you factor in the agentic loops and self-correcting cycles that vibe coding tools run continuously. For each kilowatt hour of energy a data centre consumes, it needs roughly two litres of water for cooling. A Goldman Sachs analysis from August 2025 forecasts that about 60% of the increasing electricity demand from data centres will be met by burning fossil fuels.
So when someone runs Claude in an overnight agentic loop to build the fourth Notion clone nobody asked for, they’re not just burning tokens, but burning water, electricity, and carbon… and externalising all of that cost onto everyone else while paying $200 a month for the privilege.
The outrage in the community runs in exactly the wrong direction. The question isn’t “why is Anthropic restricting my access?” The question is “why did I think unrestricted access to shared compute infrastructure was something I was entitled to?”
What ROI-Positive Looks Like
If what you’re building has a credible return on investment, the cost of tokens is irrelevant. Not small, but irrelevant. If you’re automating a process that saves your business €5,000 a month, you will happily pay €500 a month in API costs and call it a bargain. The maths works. The project is defensible. You can explain it to a client or an investor without waving your hands.
If the idea of paying market rate for compute makes the project feel unviable, then the project was always unviable. Anthropic didn’t kill it. Anthropic just made that fact legible.
This applies just as much inside organisations as it does to indie builders. I see the same pattern in companies running internal AI pilots… enthusiastic adoption of every tool available, zero measurement of what any of it produces, followed by shock when the infrastructure bill arrives. The solution is not cheaper compute. The solution is fewer, better-defined projects where the ROI case is made before the first prompt is written.
Manual first. Test the logic. Prove the value. Then invest the compute.
The Clean-Up Nobody Asked For
What’s actually happening right now, underneath the noise, is a useful market correction. Several providers of AI coding tools are revisiting their pricing strategies: Cursor changed how it priced its $20-per-month Pro plan to limit power users in mid-2025, and OpenAI proclaimed it was removing usage limits for its Claude Code competitor, Codex. OpenAI removing limits is a bold move. Watch how long that lasts before the economics force a reversal.
The developers doing legitimate, high-value work, like building real infrastructure, solving actual problems, and shipping tools people pay for, will absorb higher compute costs without complaint because the ROI is there. Albert Örwall, for example, built Moatless Tools, a framework for running coding benchmark evaluations that achieved a 70.8% solve rate at just $0.63 per instance, real work that holds up against scrutiny.
The developers building Notion clone number four, or automating a workflow they never manually tested, or racing up a token leaderboard for bragging rights? They’ll move on to wherever the next free lunch is being served. And that’s the filter doing its job.
The token tantrum will pass. Anthropic will probably improve its communication (it really should) and the limits will find a more stable equilibrium. But the underlying lesson will stick for anyone willing to learn it: AI infrastructure is not a right; it’s a resource, and treating it as if it were free has always been a fiction.
The fiction just became expensive.
That’s not a problem. That’s progress.
Leave a Reply