Claude started feeling smaller.
Not worse, exactly. More like I was doing the same kind of work and the room had less air in it. Same subscription. Same tool. Same me rambling into the prompt box.
Maybe Claude Code has a bug. Maybe I am using it badly. Maybe Anthropic changed the limit math, or the system prompts got fatter, or capacity is tight this week. The boring explanations probably cover a lot of it.
That is not the part that stuck in my head.
I hit the limit and felt the work shrink. Which is embarrassing. There are other models. Some of them are good. Some are the RC Cola of Claude, which I say with affection. But Claude understands my mess in a specific way. It catches intent before I have cleaned it up. It makes defaults I would have made after thinking for five more minutes. It lets me switch between English and Turkish brain-noise without punishing me too hard.
That is the magic for me. The input forgiveness.
Then somebody starts charging rent on it.
A little honey in the mouth
In Turkish we say ağzına bir parmak bal çalmak. There is no clean English version. The closest is: they gave us a taste of honey.
The image is weirder than the translation. Someone smears a little honey on your mouth. You do not get the jar. You get enough to know you want more.
That is what this AI moment feels like. The good version arrived first, or at least the version that felt good to users like me. The models were fast. Subscriptions were cheap enough to stop thinking about. The context windows sounded fake until they became normal, and then every app got a sparkle button like the boring parts of life had finally been deprecated.
And some of it is good. I am not doing the grumpy “AI is useless” routine. I use these tools every day. They changed how I work. That is why the economics annoy me instead of boring me.
The current era of free or heavily subsidized AI access is already being described as temporary, not stable. No surprise there, if you think about it for eight seconds. Frontier models were never going to be a public utility gifted to humanity by venture capitalists in quarter-zips.
Still, the meter feels rude when it appears.
Drug dealers figured this out before SaaS. Silicon Valley added nicer typography and a monthly plan.
We have seen this movie
Uber is the easy example because almost everyone remembers the early version.
Tap button, car appears. It felt cleaner than taxis and cheaper than it had any right to be. The word “disrupt” got thrown around like holy water.
Then the old thing died enough for the new thing to stop being cute.
Uber used venture funds to keep user costs artificially low. Later, after it and Lyft became the default in many cities, pricing got weird. “Dynamic pricing” is a beautiful phrase if your job is to make “we charge what we think you will tolerate” sound like weather.
This is not a perfect analogy. AI companies are not ride-hailing companies. A model is not a car. Fine.
The shape still bothers me.
Enshittification is Cory Doctorow’s word for platform decay: the platform wins users with the good version, shifts value away from them later, and eventually starts clawing value back from everyone. The worse version arrives after leaving gets expensive.
I do not think AI is already there. Too neat. The market is still chaotic and full of companies lighting money on fire while wearing Patagonia vests.
But the ingredients are not subtle. Workflows built around one provider. Pricing hidden behind subscriptions and rate limits. People saying “I cannot use anything else now” after eighteen months of exposure.
Fool me once, shame on you. Fool me twice, still on me.
The compute is not imaginary
AI feels like software, which is part of the trick.
You type words into a box. Words come back. No truck arrives. No machine makes noise in your garage. Nobody hands you a receipt for the GPU somewhere chewing through your request.
The IEA estimates data centers used about 415 TWh of electricity in 2024 and projects that to reach around 945 TWh by 2030. That does not mean every ChatGPT query is melting an iceberg. It means the thing has a body. The cloud is buildings and chips and water and contracts and somebody trying to make the numbers work.
So when people say AI is changing everything, I keep wanting to ask: at what price?
Not the subscription price. The later price. The one after free tiers tighten, after the model you trained your habits around moves behind a different plan, after a provider decides your kind of usage belongs in a more expensive bucket.
That is where the Claude limit thing becomes more than a personal annoyance. It is one little crack where the real economics peek through.
I do not know why my limits feel worse. I know what it rhymes with.
The money goes in circles
The other thing that makes the AI boom hard to read is the money moving around the stack.
Circular financing is the polite term. One company invests in another. The second company spends a lot of that money back on the first company’s cloud bill or chip order. The chart looks healthy. The room claps.
Some of this is normal. A supplier financing a customer is not automatically a scam. Noah Smith makes the fair argument that many AI circular deals look more like vendor finance than fake round-tripping. If Nvidia helps OpenAI buy Nvidia chips, that can be rational.
Sure.
It still makes demand harder to read from the outside. It creates a room where the same companies are investor, supplier, landlord, priest, and maybe the fire marshal too.
Amazon invested another $4 billion in Anthropic, while Anthropic named AWS its primary cloud provider and training partner. That does not mean something shady happened. It means the funding and the infrastructure are braided together before most of us can even see the rope.
Maybe that is how you build frontier AI. I can believe that. I can also believe an industry this expensive gets good at convincing itself the room is bigger than it is.
The part that makes me tired
This is where I get more annoyed than analytical.
We keep doing this. We build something genuinely powerful and immediately route it through the oldest operating system we have: money and power.
Make the thing irresistible. Let people arrange their habits around it. Later, explain that the earlier terms were never the real terms.
I do not have a clean political theory to staple onto that. I have read enough Ishmael and Sapiens-adjacent civilization critique to recognize the feeling, but I do not want to cosplay as a philosopher because my Claude Code session hit a limit. That would be embarrassing, even for a blog.
The disappointment is simpler than that. Humans keep making tools that could expand what a person can do, then the tool becomes territory. Territory gets owners. Owners want extraction. A few years later everyone stands around acting shocked that the thing got worse.
AI is not bad because it costs money. Expensive things cost money. People should get paid. Infrastructure has to exist somewhere.
The insult is being sold the subsidized version as if it were reality.
That is why the honey metaphor works for me. They put a little honey in our mouths. Now we know how good it can taste. A model understands the messy version of us and suddenly a week of work becomes an afternoon.
I am still going to use the tools. I am not moving to a cabin to handwrite shell scripts by candlelight.
But I want exit paths. I want off-brand models in the rotation. I want workflows that survive a provider having a bad quarter.
The good version was not fake. That is the annoying part.