The cheapest model is often the most expensive one.
I keep watching makers downgrade their coding agent to save money, then quietly spend more. A weaker model stalls, retries, and produces work a stronger one has to redo later. Switch models mid-session and you blow the warm cache too, so you pay full price to re-read context you already paid for once.
The unit that matters isn't the single call, it's the whole task. Usage caps and forced downgrades do lower the number, but mostly by making the output worse, which shows up as your time. Session-level cost is the thing to optimize, not per-prompt cost.
There's a real opening here for tooling that reasons about the entire session instead of the next prompt. That's the kind of gap I try to surface with SoloVault, where I look for early market signals and turn them into buildable ideas for solo makers.
How do you actually track what your agents cost per finished task, not per call?
Replies