¥15,000 a month is a lot. No point pretending otherwise.
The gut reaction is: "It's just an AI. Could we do ¥500 a month? Please? 🙏"
But the price looks different when you change what you compare it to. Don't line it up against Netflix or other apps. Line it up against the evenings and weekends you used to spend on debugging, digging through logs, remembering where you left off, and trying one fix after another.
So the real question is how many of your own hours it buys back. If the answer is "a lot," ¥15,000 stops being "AI money" and becomes a way to get your free time back.
Why "cost per question" is the wrong way to count
For a quick chat, counting questions works fine. For an AI agent that works for a long stretch and plugs into your other tools, the real cost is the legwork you used to do yourself.
That legwork looks like this:
- open GitHub
- hunt for the right file
- remember what you did last time
- read through the logs
- reproduce the bug
- rule out the possible causes one by one
- fix it
- test it
- check that nothing else broke
- come back a few hours later and load the whole problem into your head again
Each step may take only ten or twenty minutes. Repeat it enough and the whole evening is gone.
An AI earns its keep not by writing a clever paragraph, but by taking over this scattered trial and error.
The last 20% can cost more than the first 80%
On a side project, the first 80% can go shockingly fast. The screen shows up, the main flow works, the demo looks great. You think, "Basically done."
Then the long part starts.
One edge case breaks. Old data sticks around. One language looks off. An outside service throws an error. The tests pass but the live site behaves differently. Fixing A breaks B.
The first 80% takes hours. The last 20% can take days, sometimes weeks.
And the hard part isn't only the difficulty. Every time you sit down, you have to rebuild it all in your head: "Where was I? Why did I add this rule? What made the bug show up?"
That's why a long-running agent is worth the most when it can carry the whole chain: reproduce the bug, find the cause, fix it, retest, and check that nothing else broke.
An AI that makes the first 80% ten times faster is nice. One that lifts the last 20% off your evenings is worth more.
Connect it to your tools and it stops advising and starts doing
Chat alone is handy. Once the AI can reach GitHub and other tools, its job changes.
Without that connection, you are the delivery person:
look → copy → paste into the AI → read the answer → open GitHub → edit → copy the result back → explain it all again
With the connection, a lot more of it becomes:
look → compare → fix → check the real result
The AI didn't just get smarter. You stopped being the go-between.
One person's projects cross job titles: research, writing, coding, testing, sorting data. Normally you'd hire a different person for each, or do it all yourself. An AI that can work across all of that lets you slowly retire the "do-everything you."
Switching tasks is overtime you never see
Microsoft Research has reported that developers tend to call a day productive when they finish a big piece of work without interruptions or constant task-switching.[1]
So the question isn't only whether AI shrinks a 60-minute job to 50.
It's whether you can skip opening the laptop at night just to load last time's mess back into your head.
Say a bug takes 30 minutes to fix. If you first have to figure out "Where were those logs?", "What was this branch?", "Why did I add this rule?", the free time you actually lose is much bigger.
An AI that remembers the context and can act in your tools cuts that "reload" cost.
How many hours do you need to win back for ¥15,000 a month?
It boils down to one number:
effective hourly cost = monthly fee ÷ hours you got back
At ¥15,000 a month:
| Hours won back per month | Cost per hour |
|---|---|
| 5 hours | ¥3,000 |
| 10 hours | ¥1,500 |
| 20 hours | ¥750 |
If it takes one or two full workdays of evening and weekend work off your plate each month, this stops being "expensive AI" and becomes a pretty cheap way to buy time.
Of course, time spent playing with the AI isn't time saved. If you spend just as long checking what it produced, the time only moved.
DORA's 2025 research describes AI as an amplifier of whatever strengths and weaknesses a team and its process already have.[2] DORA's 2026 write-up also points out that the time saved on creating things often shifts into checking and verifying them.[3]
So buying AI isn't automatically a win. The simplest test is how many hours of your own night shifts actually disappeared.
When an expensive plan pays off, and when a cheap one doesn't
A pricey plan pays off when you:
- hand it long investigations
- connect it to your other tools
- let it run from building to checking
- stop re-explaining the same things
- give it the hunt for root causes, not just routine chores
- use it to replace the work you did at night and on weekends
- build things so one improvement helps all your future work
If you only ask a few short questions a month, a premium plan is probably a waste. And if the AI churns out piles of output that you then have to check line by line, that's risky too. DORA says AI amplifies the system you already have, so speeding up a shaky setup can add instability.[2][4]
So an AI's value doesn't come from how smart the model is alone.
The more you have in place, such as one master copy of the work, clear permissions, tests, a check of the real result, the ability to undo changes, and a way to contain failures, the more safely you can cut back on checking by hand.
Fix the setup, don't blame the worker
If a person keeps making the same mistake, "be more careful next time" doesn't fix it. The same goes for AI.
If it forgets, keep a master copy of the facts. If it stops halfway, save its progress so it can pick up again. If it says "done" too early, make "the real result checked" the finish line. If one failure stops everything, keep failures from spreading.
Rather than demanding a smarter AI, build a setup that holds up even when the AI is a bit clumsy.
The funny part: safeguards built back when AI was weaker turn out to matter more once it gets stronger. What started as protection from a clumsy helper becomes the road that lets you step out of the process.
It's not a subscription, it's the price of your free time
So whether ¥15,000 a month is expensive isn't decided by the AI itself.
It's decided by how many hours of your life it gives back.
You pay about one or two days of your day-job pay and cut the nighttime and weekend bug fixing, log digging, re-explaining, and trial and error. With the time left over, you can sleep, play, or think about something more fun.
And if the setup you build keeps saving work in later months, the savings don't stop at one month. They add up.
So yes, it's expensive. Really expensive.
But if it truly takes over your night shifts, it can be worth paying for.
Still, OpenAI, ¥500 a month would be great. 🙏
Something can be worth the money and you can still wish it cost less.
References (4)
- Microsoft Research, Software Developers’ Perceptions of Productivity (FSE 2014) microsoft.com
- DORA, State of AI-assisted Software Development 2025 dora.dev
- DORA, Balancing AI tensions: Moving from AI adoption to effective SDLC use (2026-03-10) dora.dev
- DORA, Impact of Generative AI in Software Development (updated 2026-04-13) dora.dev
