What I actually pay for AI (and what I canceled)

TL;DR: I pay for one AI subscription: Claude Max – the regular 5x tier, about $100 a month. I had other tools and subscriptions, and I canceled them, because Claude can do everything now. When I occasionally need another tool, Claude reaches it through an API at usage rates, so I pay for the amount I use instead of holding another subscription. I’m Julie Kaiser, a freelance science writer – here’s my honest math, tier by tier.

What I actually pay for

One subscription. I pay for Claude Max – the regular 5x tier – at about $100 a month. And I pay API usage for a few other tools.

If you’re not in this space, that difference is worth a sentence. A subscription is a monthly rate for a certain amount of use. API access bills you for exactly what you use. If you need a tool constantly, a subscription can be worth it; if you need it a few times a month, API pricing means those few times cost you cents, not a seat.

The subscription does the work of everything I used to pay for separately. That sentence is the entire pitch, and the rest of this post is me showing my math instead of asking you to take it on faith.

What it does on a normal week

A normal week runs through that one subscription end to end: drafting and editing this blog, TikTok videos, client work, fact-checking before anything ships, research and keyword research, editing videos, writing transcripts – and the private stuff too: telling me what to cook for dinner, reminding me when I planted the potatoes and when to dig them up, how to operate the heat pump, logging a recipe I’m testing.

None of that is exotic. It’s a freelance science writer’s ordinary week, with the personal stuff added in.

What I canceled and why

Everyone’s AI advice is “here are my 17 favorite tools.” Mine is: cancel almost all of them.

Here’s the mechanism that makes that possible. Claude Code lets me build the workflows I’d otherwise have to subscribe to. Once you can build your own workflow – and pay API access for only the amount of a thing you actually need – you don’t need a subscription for every specific job. The system that does this for me is How I built an AI that actually knows my work.

Where that replaced real subscriptions, one by one:

• Claude Code reads and writes my Notion, so I don’t need Notion AI.
• Claude Code runs Perplexity deep research for me at an API rate, so I don’t pay for a Perplexity subscription.
• One pay-per-use API account gives me occasional access to other models – my LLM council, a second opinion from a different AI without a second subscription. The same route has helped me find free, open-source transcription software to run on my own computer.
• Keyword and competitor research works the same way: I pay for the data I pull, not a monthly seat.
• I built a workflow to edit my TikTok videos, so I no longer need a CapCut subscription.
• I haven’t designed anything lately, and I’m pretty sure I can do without my Canva subscription – so I canceled it.
• And ChatGPT: now that I’m using Claude Code, I canceled it. I still have access to Codex through a client’s enterprise plan if I ever need that specific model – but I don’t have to pay for it.

So my monthly AI spend is one subscription and a few pay-as-you-go API lines, not seventeen tools. When the spend is making me money I don’t think about it. When it isn’t, I re-evaluate. That’s the whole policy, and your credit card statement will tell you in thirty seconds how far your current stack is from it.

The stuff people sell that Claude can build you

Loads of people online are selling skills and workflows for AI tools – downloadable packs that promise to do one specific job. You can buy them. What you can also do: point Claude at one as inspiration and build your own, shaped to your actual work, for free.

And be wary either way. A lot of what’s sold – or given away – is crap. Even when you download one, you’ll want Claude to read it, iterate on it, and improve it before you trust it with anything real.

The lean-stack rule isn’t about being cheap. Most of the time you’re paying for a wrapper around something your general-purpose tool can build you itself – you need to know it exists to know to ask for it.

Not always, to be fair. Some dedicated tools earn their fee – I keep the hedge because I haven’t audited them all, and neither have you. The wrapper test is the tell: ask what the tool does that your main subscription can’t, and insist on a specific answer.

Why Claude Max is worth $100 a month for me

Go back through the list above and add it up: the subscriptions I canceled, the workflows I’d otherwise rent, the API lines that cost cents. But the real math is shorter than that. When you’re a freelancer, you know very well what your time is worth – and the things Claude has automated for me easily save me that much time every month. If you’re running the same math for your own field, I’ve done it in the open for scientists and researchers and for content creation.

And if you want this math working on your own desk without the year of trial and error, that’s a thing I do with people.

And once, it paid for itself directly. Shortly after I built my wiki, a client came to me with a thought-leadership article they’d written using AI. This wasn’t AI-generated crap – these were experts in their field who had spent real time on the story, the message, and the parts that had to be in there, and then used AI to draft it. Fair. The problem: it sounded like AI wrote it. They wanted me to fix it.

I told them what fixing it by hand would cost. Above their budget. They asked if there was anything else I could offer – and I hesitantly offered the thing I’d been building. I’ve been working on a workflow, I said. It might be good enough, and it would be much faster, and so much cheaper. They took it. We did it. They liked it.

That one job paid for my Claude subscription.

My math isn’t automatically yours – I do this all day, for a living, so the tier amortizes over everything I ship. If you’re starting out, start at zero: free tiers are genuinely enough to find out whether AI fits your work at all, and if it doesn’t, you’ve learned that for nothing. The catch with free is usually your data; often you have to pay to be able to opt out of that trade.

The one rule underneath all of this

Pay for the tool that pays you back, and make everything else prove it deserves a line item. My stack got better when it got smaller. Most of those tools weren’t doing work – they were clutter.

One subscription, watched honestly. That’s my budget.