My Year of Switching AI Coding Tools 5 Times — And the $200-300 Monthly Dilemma

I've switched AI coding tools **five times** in the past year. Codex → ChatGPT (a step back) → Cursor → Google Antigravity → Claude Code (Pro → Max). Each time, I clicked "This is it" and hit the payment button, and each time, I learned something. Now, I'm spending **₩250,000-₩300,000 per month** on AI, and I'm figuring out how to make that into a side income. This post is that **honest journey** — why certain tools didn't work, what the real turning points were, and why I keep paying for expensive tools despite knowing the cost. > I'm a solo developer in Korea building AI services. I only write about experiences based on **personally paying for and using** these tools. (My tool/price evaluations are subjective, based on my own usage environment.) ## 1. One Year, Five Tool Switches — At a Glance | Timeline | Tool | Payment | Conclusion | |---|---|---|---| | 2025.05 | **Codex** | ₩200,000 | Didn't follow prompts, couldn't even find the source of errors → retreated to ChatGPT | | 2025.12 | **Cursor** | — | First "it works" feeling. OK for simple tasks, but I had to guide it for complex ones | | 2026.03 | **Google Antigravity** | ₩180,000 | Better performance but limitations in memory/agent capabilities + Claude API double-billing trap | | 2026.04 | **Claude Code (Pro)** | ₩200,000 (Additional Payment) | Top-notch problem-solving but depleted quickly → additional payment hell | | Current | **Claude Code (Max)** | ₩180,000/month | Settled. Using it well, but a fixed monthly cost | ## 2. 2025.05 — Codex, The First ₩200,000 Failure My journey with "Vibe Coding" started in May 2025, with a **₩200,000 payment for Codex**. YouTube was flooded with videos claiming "AI will write all your code." But when I actually uploaded code and gave it tasks — **for web development, it simply didn't generate what the prompts asked for.** What was even more frustrating was that I couldn't even find **where the errors were** on each page, just going in circles. The conclusion was anticlimactic: **"If I'm going to use Vibe Coding, I might as well just use ChatGPT directly."** So, I went back to ChatGPT for a while. Although my first tool was a failure, I learned something: **"Just let the AI figure it out" doesn't work.** ## 3. 2025.12 — Cursor, The First "It Works" Moment Six months later, I tried Cursor, and this time I felt a **definite improvement**. When developing at my company, **it was quite good for implementing simpler things.** It was the first time I thought, "AI coding is actually possible." However, there were clear limitations. **When delving into more complex development, I had to ask multiple times, and it was only through me guiding the direction** that we could solve the problems. As the tool got better, the realization that **"I need to set the direction"** became even clearer. ## 4. 2026.03 — Google Antigravity, and the 'Double Billing' Trap In March 2026, I switched to Google Antigravity. I moved because **its performance was clearly better than Cursor**. However, when it came to **agent-level development or game development**, issues surfaced — I had to ask multiple times, but it **couldn't properly remember past context**, and the solutions weren't clean. (This was after I had paid ₩180,000.) Here's where I found something interesting. When I switched the model to Claude within Antigravity, it solved problems properly. But after a few uses, it **quickly exceeded its limits**. So, I **connected my own Claude API** and instructed it, "If Antigravity can't solve it, use my Claude API to get it solved." But then — it was a **double-billing** structure (Antigravity subscription + my API calls). I knew the key to performance was the **model (Claude)**, but the method was inefficient. ## 5. 2026.04 — Claude Code, Through the Burnout Hell to Max So, in April 2026, I switched directly to **Claude Code**. I decided it was better to use that Claude model without double-billing. Initially, I started with **Claude Pro**, but as expected, with its good performance, it **depleted very quickly**. This led to additional payments. It felt like I hadn't used it much, but I ended up spending an **additional ₩200,000 that month**. Eventually, I ended up subscribing to **Claude Max** and have been using it **effectively** ever since. The tool changed five times, but the lesson learned was the same as before — the core of solving problems wasn't the tool's name but **"the power of the model × my ability to guide the direction."** ## 6. The Reality of the Cost — ₩250,000-₩300,000 Per Month Let me be honest and share the numbers. This is what I'm spending each month: **Claude Max ₩180,000 + GCE Cloud + Separate AI API Calls** = approximately **₩250,000-₩300,000 per month.** I have almost no income yet, but I have fixed expenses. That's why I'm constantly thinking and searching for **how to turn this into a side income.** ## 7. An Honest Conclusion — A ₩300,000 Question Looking back, though the tools changed five times, the lesson converged into one: > **The names of AI coding tools keep changing. But what solves the problem has always been the same — a powerful model × a person who guides the direction. AI is a tool to amplify skills, not replace foundation and judgment.** And that amplification comes with a realistic price tag of **₩300,000 per month.** This is my mindset: > **Do your best first (not keeping up is a bigger loss). Then, create a way to afford that cost.** This blog, the Riel chatbot, and all the development logs I share — are essentially **the process of finding an answer to that ₩300,000 question ("How can I afford to keep going?").** Can the cost of keeping up with AI be offset by the things created *with* AI? I'll continue to document that experiment.