The Cheapest Model Usually Wins Now, and That's Weird
For most of my real work, the price gap matters more than the leaderboard.
I benchmarked my daily tasks across a few models and the expensive one won maybe a third of them.
For summarizing, tagging, and drafting, a cheap model is indistinguishable to the end user.
The leaderboard says otherwise, but the leaderboard isn't my inbox at 11pm.
I now route by task: cheap for volume, expensive for the few calls where quality shows.
Paying top dollar for a classification is the new waste. The market is pricing that out as we speak.
Updated October 28, 2026: the routing rule below has saved real money since - the gap has only widened.
Frequently asked questions:
Q: Which model should I use for daily tasks?
A: Route by task. For summarizing, tagging and drafting, a cheap model is indistinguishable to the end user - the expensive one won only about a third of real-work tests.
Q: When is the expensive model worth it?
A: The few calls where quality shows: nuanced writing, hard reasoning, anything going to a client unedited.
Q: What's the waste to avoid?
A: Paying top dollar for a classification. The market is pricing that out as we speak.