The $40 Overnight API Burn — And How to Never Do It Again

I want to tell you exactly what happened — not a cleaned-up version, but the real sequence — because every single mistake I made is the kind nobody warns you about. The tutorials skip it. The documentation assumes you already know. And so you find out the hard way, at 9 cents per call, running every two minutes, all night long.

The Assumption That Started Everything

I already had Claude Pro. Twenty dollars a month. I already had ChatGPT Plus. Another twenty. I'd been using both of them for a while and I assumed — reasonably, I thought — that those subscriptions gave me API access.

They do not.

This is the thing nobody leads with. Your Claude Pro subscription and your ChatGPT Plus subscription are for the chat interfaces — the websites, the apps where you type and get a response. The API is completely separate. Its own account. Its own billing. Its own credits. You can have an active Pro subscription and zero API access at the same time. They are not connected in any way.

So I opened an OpenRouter account — OpenRouter is a single dashboard that gives you access to multiple AI models through one API key — and I added $40 in credits. That's my actual API budget. Completely separate from the Pro subscriptions.

What I Was Building

I was setting up an automated real estate lead analyzer. The goal: every time a Redfin property alert landed in my email, AI would automatically analyze it as a potential deal and push a summary straight to my Telegram. No manual work. Just a notification on my phone with the key numbers already broken down.

Here's how it was wired:

• 01
  New Gmail as the Hub I couldn't get my Outlook connected to OpenClaw directly, so I set up a dedicated Gmail account and forwarded my Redfin alerts there. Gmail connected to OpenClaw without an issue.
• 02
  OpenClaw Watching Gmail Every 5 Minutes OpenClaw — the automation platform running on my VPS — was set to check Gmail every five minutes for new Redfin emails and pass them to the AI for deal analysis, then push the result to Telegram.
• 03
  Telegram Connected and Working — Partially Telegram was receiving notifications. But instead of analyzed deal summaries, it was just getting the raw emails forwarded as links. The AI analysis step wasn't firing correctly yet. I left it running overnight to keep debugging.

The Burn

What I didn't realize was that two loops were running at the same time.

The Redfin email check every five minutes — I knew about that one. But there was a second process, something else I had configured earlier during the build, running on a two-minute loop. I don't remember exactly what it was doing at that point, but it was making API calls to OpenRouter on every cycle.

The other problem: I had OpenRouter pointed at Claude Opus — one of the most capable and most expensive models available. I didn't choose it intentionally. It was either the default configuration or something I'd set during earlier testing and forgotten to change. At the time I didn't fully understand that model selection has a dramatic impact on cost per call.

Nine cents a call doesn't sound like much. But nine cents, every two minutes, for twelve-plus hours, while you're at work and not watching — that's a lot of nine cents. I came home the next day. The entire $40 was gone.

What I Did to Fix It

I went through every scenario in OpenClaw, found all the running triggers, and shut everything down. Then I figured out exactly what went wrong at each point before I put another dollar in.

• 01
  Switch to the Right Model for the Job Opus is built for complex, heavy reasoning. Analyzing a real estate lead email is not that. I switched to Gemini 2.5 Flash — a lightweight Google model available through OpenRouter — which handles structured analysis just as well. Cost per call dropped from 9 cents down to $0.0018. Match the model to the job.
• 02
  Audit Every Loop Before Walking Away Before leaving anything running unattended, document every trigger: what it does, how often it fires, and what it calls. If you can't account for every automated process, shut it down until you can. I had a trigger running I'd completely forgotten about.
• 03
  Set a Hard Spending Limit in OpenRouter OpenRouter has a spending limit setting in the dashboard. Set a monthly hard cap. Set a soft limit that alerts you before you approach it. This alone would have stopped the burn hours before it hit zero.
• 04
  Set max_tokens on Every API Call Every call to the API should include a max_tokens parameter — a cap on how long the model's response can be. Without it, a model can return an extremely long response and charge accordingly. For lead analysis, 500 tokens is more than enough.

What This Means for You

Your Pro subscription is not API access. Budget them separately. Manage them separately. Don't assume one covers the other.

Model selection matters more than most tutorials mention. Opus and GPT-4o are the heavy lifters — built for complex reasoning, priced accordingly. For the majority of automation tasks, a lighter model handles the job just as well for a fraction of the cost. Always ask: does this task actually need the most powerful model available?

Never leave an untested loop running overnight. Watch it run for ten minutes. Keep the OpenRouter dashboard open and watch the cost per call tick up in real time. Confirm the numbers make sense before you step away.

Set your spending limits before your first dollar goes in. Not after. Before.

The automation I was building works. It's running right now. Real estate leads hitting my Gmail, analyzed by AI, pushed to Telegram. It just cost me $40 and a week of cleanup to get the configuration right. You don't have to pay that tuition.