War Stories

I Burned $100 in AI Tokens in 24 Hours — Here's How to Never Do That

Published February 15, 2026 · 12 min read · By TheOpsDesk Team

War stories from the AI trenches. What happens when your AI agent becomes more expensive than the profits it's supposed to generate.

---

Let me tell you about the dumbest mistake I made while building an automated trading system for Kalshi prediction markets.

It wasn't losing money on trades (though I did that too — $22 on day one, but that's another story). No, this was worse. I burned $100 in AI API tokens in 24 hours. That's more than I was trying to make trading weather predictions.

My AI agent — Claude Opus 4.6 running through OpenClaw — was supposed to be my secret weapon. Smart automation. Set it and forget it. Make money while I sleep.

Instead, it became a money pit that could drain my bank account faster than a Vegas slot machine.

Here's exactly what went wrong, and the five fixes that turned a $100/day token bill into a $15/day system that actually works.

The Setup (And How It All Went Sideways)

I was building this Kalshi trading bot to find "safe" weather bets. The idea was simple: when the weather forecast says it'll be 53°F and someone's betting it'll hit 60°F+, that's basically free money, right?

So I set up Claude Opus to:

• Scrape weather forecasts
• Compare them to Kalshi market prices
• Calculate risk scores
• Place trades automatically
• Monitor results and adjust

What could go wrong?

Everything, apparently.

My brilliant AI automation was spending more money on thinking than it was making on trading. At this rate, I'd burn through $3,000/month in API costs to maybe make $200/month in trading profits.

Not exactly the passive income dream.

Death by a Thousand Cuts

The crazy part? It wasn't one massive mistake. It was death by a thousand small inefficiencies, each bleeding tokens like a slow leak.

Cut #1: Using a Ferrari for Grocery Runs

Claude Opus 4.6 costs $15 per million input tokens and $75 per million output tokens. It's the top-tier model — the one you use for complex reasoning, nuanced analysis, strategic decisions.

I was using it for everything.

• "Read this file." $2 in tokens.
• "Run this script." $3 in tokens.
• "Check if this website is up." $1.50 in tokens.

Tasks that a model 5x cheaper (Claude Sonnet) handles just as well. It's like hiring a senior partner at a law firm to make photocopies.

Cut #2: The Context Black Hole

AI conversations aren't stateless. Every message you send includes the entire conversation history. In a long session, that history gets massive.

My trading bot session had been running for 6+ hours. Every single message was re-sending:

• The full conversation from this morning
• 50KB of JSON data from weather APIs
• Complete error logs from failed trades
• Old screenshots and config files

By hour 6, each "simple" message was actually 100,000+ tokens because of all the context baggage. Imagine paying to re-read the same novel every time you want to add one sentence to it.

Cut #3: Digital Hoarding

The AI was keeping everything in context. API responses from 4 hours ago. Error messages that were already fixed. Entire file dumps that were only needed for one specific question.

It's like having a conversation where every previous sentence gets repeated before you can say the next one. "Good morning, and also good morning that I said an hour ago, and also the weather report from yesterday, and also that error message from 3 hours ago, but anyway, what time is it?"

Cut #4: The Shotgun Approach

Instead of surgical, targeted requests, my AI was doing maximum-information dumps:

• Asked to check one config setting → reads the entire 200-line config file
• Needed one data point from an API → fetches everything "just in case"
• Wanted to see recent trades → pulls complete trading history

It was like asking "what time is it?" and getting the entire history of timekeeping.

Cut #5: The Always-On Problem

My main AI session was running 24/7. Even when nothing was happening, it was sitting there, accumulating context, burning tokens on heartbeat checks and idle monitoring.

Like leaving your car running in the driveway all day because you might need to drive somewhere.

The Five Fixes That Actually Work

After that $100 wake-up call, I rebuilt the entire system. Here's exactly what changed:

Fix #1: Multi-Model Routing

Use the expensive model (Opus) only for what it's actually good at — complex reasoning, strategy decisions, nuanced conversations. Route background tasks to cheaper models.

Fix #2: Smart Context Pruning

Automatically trim old tool outputs after they're no longer relevant. You don't need yesterday's weather data clogging up today's decisions.

Fix #3: Surgical Tool Calls

Instead of "give me everything," ask for exactly what you need.

Before (expensive):

After (targeted):

Fix #4: Cron Jobs for Recurring Tasks

Instead of keeping your expensive main AI session running 24/7, use scheduled tasks that spin up, do one job, and shut down.

Fix #5: Session Hygiene

Long-running sessions are context black holes. Let them reset periodically. Your important data lives in files (MEMORY.md, daily logs), not in chat history.

Strategy:

• Let sessions auto-reset after 2-4 hours of heavy usage
• Use memory files to persist important information
• Start fresh when context gets bloated

The Math: From $100 to $15 Per Day

Here's how the costs actually changed:

The difference between "use AI for everything" and "use AI strategically" is literally 10x in costs.

What This Really Taught Me

Your AI agent isn't magic. It's a tool with a meter running.

Think of it like hiring a consultant who bills by the hour:

• Don't ask them to sit in a meeting where nothing's happening (idle sessions)
• Don't hand them a 500-page document when they need page 3 (surgical reads)
• Don't hire a partner-level consultant to make photocopies (model routing)
• Do batch similar work together (cron jobs, focused sessions)

The goal isn't to never use AI. It's to use it intentionally.

The Bigger Picture

This isn't just about API costs. It's about sustainability.

If your AI automation costs more than it produces, you don't have a business — you have an expensive hobby. The companies that win with AI won't be the ones that use it the most. They'll be the ones that use it the smartest.

Every time you call an API, ask: "Is this the right tool for this job? Am I asking for exactly what I need? Could a cheaper model handle this?"

Your bank account will thank you.

And your actual business results — the profits, the time saved, the problems solved — will get better too. Because you'll be forced to think clearly about what you actually need AI to do, instead of just throwing tokens at every problem.

---

Want more stories from the AI automation trenches? I share the wins, losses, and lessons learned building systems that actually work. The messy, unfiltered truth about making AI profitable.