I run my own work plus a growing fleet of scheduled AI tasks across three machines and two separate subscription plans, and I had a clear mental model of which plan carried which machine's load. An audit found that mental model was exactly backwards.
I stopped organizing automation by which physical machine happened to host it and split it by function instead. One plan carries my own interactive work only. The other carries every scheduled and background task, moved onto the one computer that stays on around the clock.
The plan that felt like it was burning turned out to be the one quietly carrying both my live work and the entire nightly task fleet at once. An idle machine costs almost nothing. Running the wrong kind of work on the wrong host is what actually costs money. I also ruled that an expensive or careful model tier should only run on my own real-time work or genuine synthesis, never unattended in the background, because a scheduled task's mistake compounds without anyone watching it happen.
I moved the routine task fleet onto cheaper, faster models running on the always-on machine, kept my own interactive work on the stronger model, and had the fleet write its own status file so every machine sees the same picture instead of each one silently assuming the fleet's alive or dead. I also decided against adopting a second competing tool a peer group was using. It solved none of the actual constraints, it would've just added a second thing to keep running.
Don't trust your gut about which system is under load. Go verify it, because the label and the reality can be running exactly backwards.
Reorganizing the system by what it actually does, rather than by habit or an assumed label.