Over the weekend I noticed Claude made some changes to their subscription model. It got me thinking about how my usage was actually being impacted — and honestly, how little I understood about what I was paying for. The $20/month price hasn't changed — but the value out of it has quietly gone down. As AI gets more capable and does more work per session, each interaction costs more from your bucket. The flat rate stays the same; what you get out of it doesn't.
So I started looking at how it actually works under the hood.
Subscription vs. Token Bank
There are two ways to pay for AI — and understanding the difference matters for how you work. A subscription is the flat monthly rate most of us use. You pay $20/month, you get a bucket of usage, and you use it until you hit a limit. Simple. The catch is you never really know how big that bucket is or how fast you're draining it.
A token bank is different. Instead of a flat rate, you buy a balance and get charged for exactly what you use. Every word in, every word out has a cost. With Claude, that's currently $3 per million input tokens and $15 per million output tokens. A token is roughly a word, so the per-prompt cost is tiny — but it compounds fast depending on how you work.
This is where workflow really matters. On a subscription you just hit a wall and stop. On a token bank you keep going — but every long transcript you paste in, every detailed document you generate, every back-and-forth session is quietly drawing down your balance. There's no wall, just a bill at the end.
Understanding that changed how I think about my workflow.
A More Mature Question
Hitting a usage limit used to feel like a minor inconvenience. Now it's making me ask a different question — not "can AI do this" but "is this actually worth spending AI on." That shift in thinking is uncomfortable, but it's probably a more mature way to work.
The honest answer is I'm still figuring it out. I don't have a clean rule for every phase of the process yet. What I do know is that doing things manually isn't bad — it's just a different tradeoff. Some parts of the process I'm happy to own. Others I'm still experimenting with. The line isn't fixed.
Where AI Earns Its Cost
What I have landed on is prototyping. This is where AI earns its cost for me — not just because it produces a prototype faster, but because of everything that comes after it. A working prototype becomes the source of truth for the rest of delivery. From it I can generate user stories, feature maps, logic documentation, end-to-end workflows, and user test scripts. Work that could realistically take weeks gets compressed because I'm editing and refining rather than building from scratch.
That's the real value calculation. It's not about whether AI can help at any given step — it's about where the ripple effect is large enough to justify the usage.
Still figuring out where the line is — but at least I'm asking the right question now.