Models Got Cheaper, but My Bill Didn't (Again)

GPT-6.1 Sol and Gemini 4 Argon both landed at $2 per million input, cheaper than last year's flagships. Great news, I told myself, my bill will drop. It didn't. It never does, because the per-token savings get eaten by doing more tokens of more things.

The leak is structural. When a call is cheap, you stop being careful: you re-summarize, re-retry, log the whole conversation, run a second model to check the first. Each is a rounding error alone and a fat invoice together. Cheaper models didn't make me frugal; they made me wasteful at scale.

What actually cut my bill once was a usage audit, not a provider switch. I found a routine task burning frontier calls that a tiny model handled fine, and a retry loop that fired three times on every error. Fixing those two dropped the invoice more than any price cut.

The lesson: per-token price is a distraction. Token discipline is the real cost lever, and it only improves when you measure where they go. Cheaper models just hide the waste better.