Is AI an Expensive Hobby?
A real cost breakdown of building small AI apps over six months: where the money actually goes
My husband and I started experimenting with AI about six months ago. Since we were both on a career break, we were careful about spending. So we set a strict rule — no more than $100 per month on anything AI-related: APIs, tools, subscriptions, all of it.
I wasn’t trying to analyze the costs at the start. I just wanted to learn and stay within that budget. At the same time, the news about companies investing billions in AI — and developers talking about “burning tokens” — was hard to ignore. And I was building public apps, which meant other people could be burning my tokens.
Finally, I got curious: why wasn’t I running into budget issues? And where was the money actually going?
Below are the answers. It is not a deep financial analysis. It’s a simple breakdown of how much I actually spent experimenting with AI for over six months, and what surprised me.
The grand total
After six months, out of my total allowed $600, I spent slightly over $400. And here is what I spent it on:
💰 My flat-rate subscription fees were higher than my metered AI usage. I had a somewhat naive assumption that subscriptions imply a volume discount and should be cheaper than paying per API call.
Metered usage vs subscriptions
Of course, how I use ChatGPT or Cursor is very different from how I use OpenAI platform APIs. ChatGPT is for chats and architecture questions; Cursor handles all the code generation for my apps; and OpenAI APIs are for building the core logic of my apps and running them in production.
But all of them use LLMs and charge per token. So I wondered — if I paid for metered usage instead of a flat fee subscription during those 6 months, where would I end up?
I downloaded all the logs from Cursor and ChatGPT and got the overall amount of tokens, images, files processed, tools used, etc. Then I applied the rates from the latest GPT-5.4 model to see if I was under or overpaying.
GPT-5.4 pricing:
- Input: $2.50 / 1M tokens
- Output: $15 / 1M tokens
- Images: $0.17 per image (estimated)
- Uploaded image processing: $0.015
- Web search: $0.01 per call
💰Ouch! I am overpaying 2x for Cursor and 6x for ChatGPT compared to what I would have paid with metered usage. But these tools don’t offer a pay-per-use option. Also, subscriptions made it easier to work without interruptions or hitting limits. Whether that’s worth the price is still an open question.
Development vs Production cost
Another insight I got right away is that the development cost of AI apps is way higher than running my apps in production. Here is another way to look at the same data:
Both ChatGPT and Cursor subscriptions are pure development cost, Render is pure production cost, but OpenAI is a mixture of both. To split this line into production vs development costs I needed a deep dive into OpenAI API logs and billing data.
After 6 months I ended up with two small apps:
- A small text adventure game based on The Lord of the Rings (LOTR)
This one uses OpenAI APIs to generate a different journey depending on the player’s input. All I use here is the Responses API. I played a little bit with images, but quickly abandoned the idea because they were too expensive. - Virtual Alexandra (VA)
A chatbot that answers users’ questions based on my writing and instructions using APIs and retrieval augmented generation (RAG) from OpenAI. I also tried fine-tuning (didn’t work) and did a bunch of data preparation for RAG, including translating several MB of text from Russian into English and vice versa.
I used different API keys for production and development of both apps, so I could see how much I spent on each of them. Images and fine-tuning are not used in production, but they are still part of the development cost — you pay even if you fail.
💰Most of my AI spend went into development, not production. The production cost is about 10 percent of the total ($41 out of $422).
Apparently, my fear of public apps becoming a money sinkhole didn’t materialize. Partially this is because of low usage — I had about a hundred people try my apps. And these are not productivity apps people would use every day, there is no stickiness in them.
But another reason is that I was more careful with what I put in production. I optimized system prompts and discarded expensive features like images. One question that Virtual Alexandra answers costs me $0.088 — less than ten cents. Ten steps in the LOTR game cost $0.037 — about four cents.
Local models vs paid APIs
$600 was our family total AI budget, and I spent more than $400. How much did my husband spend? Well, the answer is … $0. Our interests led us into different experiments: I wanted something I could share with other people, while he was happy exploring how he can improve his day-to-day coding and test the power of the local models. But that zero is deceiving. Here is what actually happened.
💰I paid for convenience and speed, while my husband Dmitry paid upfront in hardware and setup effort.
We didn’t buy new laptops, we used what we had in the house. If you happen to have a powerful computer already and want to start your vibe coding journey, you can spend exactly $0. If your computer is from Apple, read Dmitry’s guide on setting up local AI on a Mac to get started.
But if you only have an old MacBook Air, you can still go ahead — it will be way cheaper to start experimenting with paid subscriptions and APIs instead of investing in new hardware upfront.
Learnings
AI is an affordable hobby
All my data is related to small scale apps and experiments, it doesn’t apply to corporate, startup, or deep research worlds. But for a casual hobby, just set your budget. Without watching tokens or hours and never hitting any hard limits, I was able to keep it under $100 a month.
At small scale, the true cost is app development, not running in production
Public APIs turned out to be cheaper than I expected. Unless you build a really viral app, you are very unlikely to burn a crazy amount of money. The two apps I created each cost less than ten cents per user session, so I can afford hundreds of daily users without exceeding my monthly budget.
However, development costs add up quickly. Experimenting with features like fine-tuning and image generation can easily dominate the total spend, even when production usage stays low.
AI subscriptions provided less value than I would like
It’s worth looking at your AI subscriptions — are you maximizing their value? And if not, are you paying for something else like a good uninterrupted workflow? One of my personal takeaways from this write-up is to dig deeper into this and see how much I really need these services and whether I can replace them with local AI setup.
Originally published at http://adandai.wordpress.com on April 23, 2026.
Is AI an expensive hobby? was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.