I'm learning AI fast building things as a non-technical person. My latest learning is about token/API spend for LLMs.
Openclaw just ate through $10 of API credits while I was asleep, running everything through the expensive flagship GPT 5 models, yikes. I maxed out my Codex usage limit yesterday (on day 1 of the week) and I'd always planned on moving off my ChatGPT/Codex subscription to fully use Openclaw via OpenAI APIs so I did it yesterday. Little did I know it was "move fast and go broke" vibes, and I didn't really pay attention to what model the API would use, and it stayed the same as what it was using through Codex. Ten dollars of overnight usage later, I'm needing to make better decisions about models, use cases, routing, cost management.
Thankfully there people and their AIs thinking about this and building tools to solve it, so I'm trying one of them out. Manifest routes requests based on use cases/needs/complexity to the best, most cost-effective model. As an aside, this is all vibe-code on vibe-code. I'm vibe-coding (random apps and integrations), using a vibe-coded AI gateway (Openclaw) using a vibe-coded AI router (Manifest). My fiancee had me sign up for a 3-day trial of a vibe-coded app to find out my "dosha" (?), and someone out there's making $15 a month from a bunch of people who forgot to cancel their trials (I have to think that's the business model). We live in weird times, man.
Of course the other thing that will help is not sending my AI stupid, pointless messages out of the blue just to see what it says back to me. That one's on me 🤦♂️
Openclaw just ate through $10 of API credits while I was asleep, running everything through the expensive flagship GPT 5 models, yikes. I maxed out my Codex usage limit yesterday (on day 1 of the week) and I'd always planned on moving off my ChatGPT/Codex subscription to fully use Openclaw via OpenAI APIs so I did it yesterday. Little did I know it was "move fast and go broke" vibes, and I didn't really pay attention to what model the API would use, and it stayed the same as what it was using through Codex. Ten dollars of overnight usage later, I'm needing to make better decisions about models, use cases, routing, cost management.
Thankfully there people and their AIs thinking about this and building tools to solve it, so I'm trying one of them out. Manifest routes requests based on use cases/needs/complexity to the best, most cost-effective model. As an aside, this is all vibe-code on vibe-code. I'm vibe-coding (random apps and integrations), using a vibe-coded AI gateway (Openclaw) using a vibe-coded AI router (Manifest). My fiancee had me sign up for a 3-day trial of a vibe-coded app to find out my "dosha" (?), and someone out there's making $15 a month from a bunch of people who forgot to cancel their trials (I have to think that's the business model). We live in weird times, man.
Of course the other thing that will help is not sending my AI stupid, pointless messages out of the blue just to see what it says back to me. That one's on me 🤦♂️