It happened on a Friday night.
I was in the middle of building a new tool — a social icebreaker game called the Loneliness Test. I had a working prototype but it felt off. The invite-copy feature wasn't right. The animations were janky. I kept asking Claude to iterate.
Then this appeared in my terminal:
Current week (all models): 100% used.
Not a crash. Not an error. Just a soft wall. Resets Tuesday night.
The dashboard told the full story. $36.43 burned in a single 4-hour session. 83% came from subagent-heavy sessions — background workers scanning files, iterating on code, making changes across the project. By Friday night, the week's quota was gone.
I had three days until the reset. The tool was half-finished. I refused to wait.
The Geeky Workaround
I opened Gemini and explained my situation. The advice sounded clever: bypass Claude's limits entirely by using Aider — an open-source terminal coding assistant — wired to DeepSeek's API, which was supposed to offer "95% of Claude's coding ability at one-tenth the price."
Within twenty minutes I had it running. I typed a test message: "Hello."
Tokens: 10k sent, 69 received. Cost: $0.0028 session.
Aider had silently read my entire repository — over 1,000 files — compressed the architecture into a Repo-map, and sent it all to DeepSeek before I'd said anything useful. Nearly three cents to say hello. Fine. Let's get to work.
I gave it a simple task: clean up the invite-copy feature in my HTML file. Fix some text. Update a few strings.
What followed was forty minutes I will not get back.
DeepSeek couldn't find an exact text match because of a whitespace difference between what it expected and what was actually in the file. Instead of adjusting its approach, it panicked. It kept modifying its search slightly, failing, then asking:
"Try this? (Y/n)."
Y.
"Try this? (Y/n)."
Y.
"Try this? (Y/n)."
I pressed Y seventeen times. The file was completely untouched. I hit Ctrl+C and checked my DeepSeek billing.
¥0.49. Seven US cents. Forty minutes of watching a machine argue with HTML whitespace, achieving nothing. The total DeepSeek bill for all of June was ¥12.27 — about $1.70. Cheap. But I had nothing to show for it.
The Gemini Marathon: 30+ Versions in One Weekend
I gave up on the terminal agent entirely and switched to Gemini's chat interface.
This actually worked — Gemini could hold context, understand what I was describing, and generate complete HTML files from scratch. The problem was the iteration cycle:
Describe what's wrong → Gemini generates new file → download it → open in browser → test → something's still off → repeat.
No version control. No memory between sessions. Just me, a growing folder of identically-named HTML files, and a browser history I'd rather forget.
That screenshot is from one twenty-two-minute window. There were more before it and more after it. By the time I was satisfied, I had gone through more than thirty versions of the same file.
I also pulled in ChatGPT at one point — not to write code, but to review what Gemini had produced. Running one AI's output through a second AI for critique turned out to be genuinely useful. ChatGPT spotted a logic bug I had missed. Gemini fixed it in the next version.
It took a full weekend.
What Actually Came Out of It
The Loneliness Test has thirteen levels of solo activities — from shopping alone (Level 1) to going to surgery alone (Level 13). At each level you answer honestly: Totally fine or I can't do this. When you stop, it generates a personalized message you can send to a friend to invite them out.
After thirty-plus iterations across three different AIs, the final version is genuinely good.
I could not have built it without AI. It just wasn't the AI I expected to use.
The Irony I Noticed at 1am
Somewhere around the fifteenth Gemini download, I noticed something.
I was building a tool about loneliness — about the discomfort of doing things alone — while experiencing the loneliest coding session I'd had in months.
Claude is the AI I've worked with long enough to have a shorthand with. I describe something half-finished, it understands what I mean. I say "the animation feels off," it knows which animation. There's a fluency that builds up over time, over a project, over shared context.
Gemini didn't have that context. DeepSeek didn't understand my file structure. ChatGPT hadn't seen the project before. I had to re-explain everything from scratch every thirty minutes.
Losing access to your main AI tool doesn't feel like losing software. It feels like your experienced colleague went home for the weekend and left you with people who need the entire project explained from the beginning.
The Honest Verdict
DeepSeek is great for reasoning about code. It's unreliable for autonomously executing changes in a real project. The raw intelligence is impressive. The tool-calling — navigating a file system, running commands, overwriting code without breaking things — is not there yet. When it gets stuck, it doesn't recover gracefully. It asks you for permission to fail in a slightly different way.
Gemini's chat interface works, but iterating without persistent memory is exhausting. You become the memory layer. If your project has any real complexity, you spend as much time re-explaining context as you do describing what to fix.
The $20/month Claude subscription is genuinely one of the best value propositions in AI right now — not because it's cheap, but because the moment you lose it, you understand what you were actually paying for. The continuity. The context. The ability to say "you know that file we were working on" and have that mean something.
I am not sponsored by Anthropic. I'm just someone who found out the hard way.
Try the Tool
If you want to see what thirty-plus iterations of AI output and one weekend of token anxiety produced: The Loneliness Test — free, no login, works on mobile.
Play through the thirteen levels. See where you stop.
What level did you reach?
Top comments (0)