You Probably Don't Need a Frontier Model. Here's the Math.
Frontier pricing only makes sense if the task pays for it. For most work, a smaller model at a tenth the cost wins.
The frontier models are astonishing and, for most of what I do, a waste of money. GPT-6.1 Sol at a fifth of Astra's price, Claude Sonnet 5.5 at 30% lower cost per task, and a dozen open-weight models you can host yourself all handle the actual jobs: drafting, summarizing, classifying, light coding. The flagship's edge rarely shows up in the work that pays my bills.
Do the math on a real task. If a smaller model gets you 95% of the result at a tenth of the cost, and you run it ten thousand times a month, the frontier model is costing you nine times more for a sliver of quality nobody notices. The 5% is real on a hard reasoning problem and invisible on a support reply.
Where the frontier still earns it is the genuinely hard stuff: long-horizon coding, messy multi-step research, anything where being wrong is expensive. I keep one frontier subscription for those, and route everything else to the cheap tier. The split is the whole optimization.
The trap is defaulting to the best because it's the headline. Match the model to the task's stakes and volume. Most of your AI spend should be on the model nobody writes about, and that's fine.