The AI Subscription Creep: How We Went From $20 Chatbots To $400 AI Plans

Grok just launched a $400/month AI plan. That is not a typo. Four hundred dollars, every month, for access to an AI chatbot with some extra features. If that sounds insane, here is the part that should worry you more: a lot of people are already paying close to that across multiple AI subscriptions without realizing it.​If you are paying more than $60/month total for AI tools right now, you are caught in something the industry quietly built over the last three years. This piece is about how we got here: the pricing experiments you funded without knowing it, and why your AI bill feels like it appeared out of nowhere. This first piece is about how we got here: the messy experiments, the token math behind the scenes, and how the industry slowly collapsed into a small set of pricing models that most people never asked for.

How generative AI really charges you

Before getting into pricing strategies, I want to zoom in on how a typical generative model works from a cost perspective. When you type a prompt, the system does three things:It turns your text into tokens in.It runs a bunch of internal computations.It turns its answer back into tokens out.Tokens are just chunks of text. The lab is not really counting “questions”, it is counting those chunks. That is what drives their real cost.For a long time, the mental model was simple. You send some tokens in, the system processes them once, and you get tokens out. With newer reasoning or “thinking” models, there is an extra internal loop where the model generates intermediate thoughts for itself, which are also tokens that cost money to compute. You never see those tokens, but the lab still pays for them.​So on the lab side, the cost model is always token based, even if the website is telling you about prompts, credits, or “Pro” limits.

The great pricing experiment

When generative AI went mainstream in 2022, nobody really knew how to charge for it, beyond the old SaaS playbook. What followed was a three year pricing experiment that you could almost treat like a lab study in public.

✦ Token based: the developer default

The first stable model was raw token based pricing, mostly exposed through APIs. You pay for input tokens plus output tokens, often in batches of thousands or millions.This setup is very fair on paper. If you send a tiny request, you pay cents. If you send a giant document and ask for a long answer, you pay more. Many pricing reports now describe this as usage based or consumption based pricing, and it is becoming the backbone of serious AI.The catch is that regular users do not think in tokens. If a tool tells you that you used 420,000 tokens today, that number does not map to any feeling of value or progress.

✦ Prompt based: simple in theory, broken in practice

To make things feel more “human”, some early tools tried prompt based pricing. Instead of charging for tokens, they said something like: pay 20 dollars and get 50 prompts. A prompt here meant one full question plus one answer.On paper, that is easier to understand. The problem is that a “prompt” can be tiny or enormous. Once people realized this, they started doing prompt stacking. They would stuff very large inputs into a single prompt and ask the model to do a lot of work in one turn.The underlying cost to the lab was still token based, so one crafty stacked prompt could be worth many normal prompts from a cost perspective. That made the economics hard to sustain at consumer friendly prices. You saw some of these tools quietly move away from strict “X prompts for Y dollars” models once they ran into this wall.​

After prompt based pricing started to break, many AI tools moved to pure subscription language. Instead of telling you how many tokens or prompts you were buying, they told you that for 20 dollars a month you got “higher limits”, “priority access”, or “Pro usage”.Behind the scenes, the lab still had to tie each tier to some internal token limit, often with a margin on top to ensure profit. For example, if 20 dollars worth of raw tokens might cover two million tokens of usage, they might design the plan so that a typical user can effectively get around one million tokens before hitting some soft or hard limit.You never see that number.

You just see that sometimes the model slows down, or refuses longer requests, or nudges you to upgrade. That opacity is part of what makes people feel like subscriptions are slipping out of their control.​On top of that, some companies added “credits” as yet another abstraction layer. Instead of talking about tokens, they sold you 10,000 or 50,000 credits.The issue here is that a credit is rarely defined in a way that you can map back to tokens. Twenty thousand credits might roughly equal ten thousand tokens. It might be some weighted mix of image generations, audio calls, and text calls. You do not know unless they publish the mapping, and often they do not.​Credits give the lab maximum freedom to change the economics without you noticing, which again feeds the sense that AI pricing is something you are supposed to accept, not understand.

How we got to where we are right now

By 2025 and into early 2026, the chaos started to settle. If you strip the branding away, most AI pricing now falls into two buckets.

1. Usage based (token or call based)

On the infrastructure and API side, usage based pricing is winning. Reports on AI and SaaS pricing describe a strong shift toward pay as you go models, especially among faster growing companies. These companies charge per million tokens, per thousand requests, or some other clear usage metric.This model lines up with the real cost structure. If a request uses more compute and more tokens, it costs more. If a model is much cheaper to run, that shows up directly in a lower price per million tokens.​It is still intimidating for non technical users, but for teams that build with AI at scale, this is the model that makes sense.

2. Subscription with soft limits

On the consumer and prosumer side, subscription is still the default. Most AI tools present a simple ladder: free tier, basic monthly plan, maybe a higher “pro” or “max” plan that promises more power.Under the hood, many of these subscriptions are hybrid. There is a fixed monthly fee, but each tier hides one or more internal usage ceilings, often still measured in tokens or calls. Once you cross those, the product rate limits you, downgrades quality, or suggests you upgrade.​Pricing surveys show that while almost all traditional software still leans on pure subscription, AI features are already shifting toward blended subscription and usage mixes. That is the coalescing you were pointing at. The surface experience is “pay monthly”. The underlying economics are “pay for what you use”, just with more blur.

Then there's the part nobody usually notices: mental overhead

So far this sounds like a business story. Labs experiment, some models win, others die. The missing piece is what this has done to people on the other side of the screen. As more AI tools launched, most people did not sit down and design an intentional stack. They did what made sense in the moment. Add ChatGPT Plus for general work. Add a coding assistant that promises better completions. Add an image tool for thumbnails. Add a research assistant. None of these look outrageous on their own. By 2025, a lot of people woke up to the fact that they were running five or more subscriptions at once. Articles on “subscription creep” and “the great unsubscribe” started to show up, and not just about Netflix and Spotify. In creative fields in particular, 2026 is already being described as a breaking point for subscription fatigue.The pain is not only financial. There is a mental burden baked into this pattern:You have to remember which tool is good at what. You have to keep track of renewals, free trials, and upgrade offers. You feel guilty if you do not “use it enough” to justify the cost.You worry that canceling will lock you out of a feature you will need later. Small business pieces now warn that this kind of AI subscription creep can quietly wreck a budget, especially when each tool feels like “just another 20 dollars”. That same creep also eats attention and energy, which is exactly what these tools were supposed to free up.​

In conclusion....

This first part is the map. We went from cheap, simple subscriptions to a mix of usage based and subscription models, with a lot of confusion layered in between. Along the way, both prices and mental overhead climbed. In the second part, I want to stay very practical. We will look at how to use token based pricing without needing a PhD, how to spot when a subscription is giving you less than you think, where the rare “good deals” actually live, and how to design an AI stack that does not quietly turn into a full time job to manage.