Can you actually go broke on AI subscriptions? I pay over 44,000 yen a month, so I looked up what everyone else pays

Oh no. I am spending this much on AI.

That was my first reaction to actually adding up what I pay.

I had been deliberately not looking.

100 dollars to ChatGPT. 100 dollars to Claude. 60 to Cursor. Plus Gemini and Notta. Cloudflare is free, and Codex and Claude Code are included in other subscriptions.

The total is 260 dollars plus 5,408 yen a month.

At a nominal 150 yen to the dollar, that's 44,408 yen a month. 532,896 yen a year. On top of that, Superwhisper, bought outright for 249.99 dollars.

...So, AI subscription poverty. Is that a real risk?

Hanageruge's equipment list

First, what I currently pay for.

Laid out like that, it's quite the loadout.

And ChatGPT, Claude and Gemini overlap on "research" and "writing". Claude Code, Codex and Cursor are all, broadly, development tools.

Honestly, I can't claim there's zero waste.

So how much does everyone else pay?

For most people, "normal" is free, or one service

The first thing I found: the average depends entirely on who you ask.

PNC Bank analysed credit and debit card transactions and found that as of May 2026, 2.2% of households were paying for generative AI. Among those households, the average spend was 31 dollars a month. Across all households, "I pay nothing for AI" is still overwhelmingly the majority (PNC Consumer Health Check).

Meanwhile, Goodwater surveyed only US consumers who use AI: 46% paid for at least one, and 22% subscribed to several. Among daily users, the paying share rises to 57% (Goodwater 2026 US Consumer Survey).

"Only 2.2% pay" and "46% pay" are very different claims.

Neither survey is wrong: the first is bank customers' transactions, the second is a survey of AI users. Across the population, most people don't pay. Among people who use AI daily, paying is no longer unusual.

Even Japanese IT engineers average 5,776 yen a month

In Japan, Findy surveyed 605 IT and web engineers.

47.1% paid for AI tools personally, at an average of 5,776 yen a month. The highest paid-subscription rates were for ChatGPT, Claude and Cursor (Findy survey).

Another survey of 375 IT engineers found 57.8% paying, with the most common band at 2,000 to 4,000 yen a month, and 18.9% paying 8,000 yen or more (Freelance Board survey).

My 44,408 yen a month is about 7.7 times the Findy average.

My figure is converted at a nominal 150 yen to the dollar. Survey populations and methods differ, so this isn't a strict like-for-like comparison.

On the numbers alone, I am a heavy user, full stop.

That said, stacked high-tier plans are increasingly common among developers. YipitData's analysis of e-receipts found 21.4% of Cursor subscribers also paying 100 to 200 dollars a month for Claude Max, and 34.6% also on Claude Pro at 20 dollars (YipitData).

Not typical. But in the world of people using AI heavily for development and production, I'm not a mutation either.

What are people paying for?

Deloitte's US survey lists the main uses of generative AI as writing and editing text, research, summarising information, explaining difficult material, and note-taking.

Among free users, the reasons for not paying were "the free version is enough" at 50%, "I don't use it often" at 20%, and price at 17% (Deloitte 2025 Connected Consumer).

So what separates free from paid isn't only which AI is smartest.

Is there work you repeat every week? Are the free tier's caps and features actually getting in your way?

That looks like the deciding factor.

Some people on X found they'd bought too much plan

How do actual users feel? I looked at X too.

Software developer Jorge Castillo dropped the 200-dollar Cursor plan for Claude Code at 100 dollars, and posted that after a month he had never once hit a limit, so he may have been paying more than he needed (post on X).

The scarier direction is usage-based billing: extra credits, APIs.

Marc-André Moreau posted about spending 120 dollars in 24 hours on Cursor, with a spike of 57.3 million tokens in a few minutes. He wrote that he couldn't tell what had caused the surge (post on X).

With a fixed monthly fee, you know your ceiling. With usage-based billing, reading long documents or large amounts of code behind the scenes can run the number up quickly.

If AI subscription poverty happens, it doesn't happen the moment you sign up for one 20-dollar plan. It happens like this:

• Stacking up AIs with overlapping purposes and never cancelling
• Upgrading to a higher tier "just in case", before you ever hit a limit
• Not setting caps or alerts on usage-based billing
• Continuing to pay because it "feels useful", without checking what got finished

That accumulation is the dangerous part.

I'm not buying AI, I'm borrowing hands

There is a reason I pay this much.

I'm not an engineer.

And yet I develop this blog myself, design it, write the posts, make the images and video, and publish it. On top of that: meeting records, organising documents, research, study, and endless small daily tasks.

Normally, each of those is work you'd ask a specialist to do.

• Ask a video editor to organise footage and cut it
• Talk structure and prose over with a writer or editor
• Ask an engineer to build and fix the blog
• Ask an assistant to organise meetings and records
• Ask a teacher the same question as many times as you need

Paying someone by the hour. Outsourcing. Bringing in an assistant or a tutor.

My AI spending feels close to that.

Stuck on a sentence, I talk it through. A word I don't know, I ask on the spot. A long document, we read together. Information I can't hold in my head gets recorded and shaped into something searchable later.

Hire one person and they have strengths and weaknesses. AI is the same, so I split the roles: Claude for talking about writing, Claude Code and Codex for implementation, Cursor for development while I work, Notta for transcription, Superwhisper for long spoken instructions.

Which is how the subscription count grew.

AI is not an actual employee, obviously.

It makes mistakes. It takes no responsibility. Without instruction it doesn't notice work that needs doing. Especially for accounting and anything published externally, I or a specialist has to check at the end.

Even so, having something beside you that moves, thinks and records with you means work that would have stalled keeps moving.

It's covering my stamina and my brain too

I'm not the type who can push through everything on energy alone.

The stamina, focus and memory I have in a day are all finite. However much I want to do, doing the research, remembering it, building it and checking it entirely alone stalls somewhere in the middle.

AI doesn't only help with working hours.

It holds half-formed thinking. It finds the part I need in a long conversation or document. It organises what to do next. It re-explains what I've forgotten.

Not having to carry it all in my own head is, by itself, worth a lot to me.

How many hours of help does 44,000 yen a month need to buy?

So how much use makes this pay for itself?

At 44,408 yen a month, if you value your hour at 5,000 yen, that's about 8.9 hours a month. At 3,000 yen, about 14.8.

This isn't "hours with an AI window open".

It has to be measured in actual output: work I didn't have to outsource, posts and features that reached completion, research time that got shorter.

Conversely, however many hours you spend with AI, if you were only playing with comparisons and nothing got finished, nothing has been recovered.

This "is it paying off at work?" view comes back in Part 2, as "if you use it for work, look at ROI rather than a ratio to income". Paying as a hobby and paying for a work tool need different rulers.

So is my usage crazy?

On the amount alone, fairly crazy.

About 7.7 times the Japanese IT and web engineer average. Several 20-dollar plans stacked on top of each other.

Ask whether the usage is crazy, though, and I don't think it is.

I don't subscribe in order to be a person who knows about AI.

I edit video. I write. I research. I study. I sort out accounting work. I make content. And, not being an engineer, I develop and publish this blog.

I'm borrowing the hands needed to do all that alone.

And right now I'm in the middle of building the structure of this blog and my work. Assembling the frame from nothing is exactly the phase that eats the most AI.

Once the foundation exists, I may not need to use it this hard every month. At that point I can drop tiers, or cancel the services whose roles overlap.

Signing up for an expensive plan isn't a vow you have to keep forever.

Use it properly while you need it. Review it once the frame is up.

What keeps you out of AI subscription poverty isn't picking the cheapest plan at any cost. It's being able to explain, yourself, what you're paying for help with.

Honestly, if you ask whether spending this much has made me rich, the answer is no.

It has got me to the point where ordinary life isn't a problem. And it has created a fair amount of spare time.

Except that with the spare time, the number of things I haven't done has, if anything, gone up.

That sounds contradictory.

Thanks to AI, I can now start work I would previously have given up on. Writing, video, research, development, all moving. As what I can do expanded, so did what there is to do.

Even with AI doing the work, my own capacity is finite. Checking. Prioritising. Deciding what to drop. Choosing which way to go next.

The more you can build, the more decisions there are. That part can't be handed over.

Which way to steer is, in the end, a human call.

Probably, once the frame of this blog and my work is in place, this phase of hammering AI will settle down. Then I'll review the subscriptions again.

My AI spending is not typical.

But for me right now, it's the cost of supplementing my stamina and my brain, and moving work forward that I couldn't do alone.

That's Hanageruge, spending this much on AI and somehow not broke.

Next: so how much is "normal"?

Writing this, one thing became clear.

"7.7 times the average" is a fact. But whether 44,000 yen a month is a lot can't be answered without a ratio to income. 40,000 yen for someone taking home 300,000 and for someone taking home a million are entirely different weights on a household. Same as rent.

So next time, using general household guidelines like rent benchmarks and the 50/30/20 rule, I calculate what percentage of take-home pay counts as normal for AI. There's a checker in the middle of that post so you can judge your own on the spot.

• Part 1: Can you actually go broke on AI subscriptions? (this post)
• Part 2: Is AI subscription poverty real? Calculating a spending line from household benchmarks, with a checker
• Part 3: Is that AI spending actually expensive for your life? (building a standard household from official statistics)

---

Prices and survey data were checked on 31 August 2026. The actual amount paid for dollar-denominated services varies with exchange rates, consumption tax and card issuers' overseas handling fees. Annual plans are billed in one payment. "No extra charge" for Claude Code and Codex means they're included in the respective plan's allowance, not that usage is unlimited.

Accounts on X and forums are self-reported by the posters and do not represent typical spending.