The $100k bill from one leaked API key

When I started running AI workloads in production, the conversation in my head was always about model capability. Which model handles this task best, which one is fastest, which one fits the context window. It took a leaked key scare to shift my thinking to a different question.

What's my financial exposure if this key gets compromised?

Some frontier providers charge per token without a hard spending cap. If a key leaks, you're on the hook for whatever gets consumed before you catch it. One bad commit, one misconfigured env variable, and the liability is open ended.

That's what pushed me to OpenRouter. Not the model selection, though that's useful. The controls. I can lock an API key to specific models so a compromised key can't call anything expensive. I can set monthly and total spending caps that actually stop usage when they're hit. The worst case becomes a number you chose, not a surprise.

The tradeoffs are real. Another proxy layer means another dependency and a few extra milliseconds of latency. Model availability depends on their routing. And the control I miss most from my previous provider is IP restrictions. Locking a key to specific IPs hopefully meant even a leaked key was useless from an unauthorized network. OpenRouter doesn't support that yet.

I still run an RTX 3090 with Ollama for batch processing and BI reports where local models handle the work fine. But for the workloads that need frontier models in production, the provider choice became a risk management decision as much as a capability one.

I don't think enough people are evaluating API providers by their financial controls. Model quality matters. So does what happens when something goes wrong. If you can not answer off the top of your head what API Key restrictions are in place, this is something you need to figure out in 24 hours.