I’ve had nights where my API bill moved faster than my cursor. It starts with a normal request, then a small correction, then another. By 3 a.m. I’m clicking Regenerate and telling myself the next response will be the one. The model isn’t a casino dealer, but the rhythm can feel similar: same action, unpredictable reward, and a meter running in the background.
Same Action, Random Reward
Game designers call this variable rewards: you repeat one action and get a random outcome—nothing, a small win, or a rare jackpot. AI chats can work the same way. One prompt returns clean code on the first try, and I feel like a prompt wizard. The next ten replies invent variables, forget constraints, or lose the thread. The early jackpot is what keeps me pulling the lever.
Paying to Iterate, Not to Finish
Providers don’t only charge for the final answer; they charge for the process. Every clarification adds to the context. As the thread grows, each new request carries more input tokens, and the conversation gets more expensive to continue. A long back-and-forth can cost more than starting over with a shorter, clearer prompt. The painful part is that the thread often gets worse: after several failed fixes, the model starts treating its own mistakes as context.
The cost curve isn’t linear. Early messages are cheap. Past a certain point—when the model starts drifting on its own errors—costs and confusion both climb.
cost ^ / | / | / | ______/ | ____/ | ___/ +------------------------> messages cheap drift regret The Sunk-Cost Trap
The gambler’s rule shows up in chat: 'I already spent $3 and mapped the architecture. If I close now, that work is wasted.' So I send fifteen more requests to prove a model wrong, instead of resetting or opening an editor. Sometimes five minutes of manual work would beat another ten dollars of regeneration. The house wins because the house controls the context window.
The Rule of 3
Here is the hardest boundary I’ve set:
prompt -> bug -> clarify -> bug -> STOP No third attempt in the same thread. Close the session. After 3–4 failed fixes, the dialogue turns into mush and the model loops on its own bad output. Starting over with a shorter prompt usually beats another round of 'sorry, here is the corrected version.'
Local Drafts and Model Tiers
Not every request deserves a flagship API. If I need to format JSON, parse a log line, or draft boilerplate, I try a local model first. Qwen or Llama through Ollama or vLLM runs on my own hardware, costs nothing per token, and keeps data private.
Model tiering helps too:
- Tier 1: local or free models for drafts, parsing, and simple refactors.
- Tier 2: cheap, fast APIs for routine bulk work and first-pass tests.
- Tier 3: heavy flagships only when I’m stuck on a hard architecture problem.
Don’t use a cannon on a sparrow.
Break Prompts Into Atomic Pieces
Long prompt novels are another way to gamble. I used to write a page of context and hope the model would build a spaceship in one shot. Now I split tasks: one piece, one clean prompt, one strict result. It is easier to verify and cheaper to retry because there is less context to carry. If a task needs ten steps, ten small prompts can be clearer than one giant one.
This article was originally published on AI Release — https://ai-release.org. Shorter notes and updates live in the channel: https://t.me/ai_release1.
Takeaway
AI is a tool like a compiler or a soldering iron, not a slot machine that owes me a jackpot. When I force it into a pipeline with hard limits—stop before the third attempt, use local models for small jobs, escalate only when needed—the dopamine loop breaks and my card balance survives. The model doesn’t need another regenerate; sometimes I just need to close the tab.

