A strong model, even on its lightest thinking mode, almost always beats a weaker model thinking as hard as it possibly can. The reason: the model's tier sets its capability ceiling, and the thinking mode only decides how much of that ceiling it actually uses. Extra thinking doesn't add new knowledge, it just lets the model dig deeper into what it already knows.
There's one exception: tasks with a clear logical path, like math, debugging, or multi-step planning. There, the extra thinking helps the model catch its own mistakes along the way (that's real inference-time scaling, it just only works there). But on every other task, the ones that need nuance and judgment, a weak model trying too hard just writes you a beautifully worded wrong answer.
I own the token budget on our team, and this is exactly the question that came up this week in our alumni group, in a phrasing I stole on the spot: head of the lions, or tail of the foxes. From there the cost question follows immediately, and here's the part most people miss: heavy thinking mode always burns a lot more tokens, because all that thinking is priced as output. And output, across every tier, costs five times more than input. So a strong model on measured effort comes out cheaper than a weak model on ultra, as long as it uses less than five times the tokens, and that's almost always what happens. A task where the strong model wraps up in 500 output tokens costs us about 1.25 cents at our top tier. The exact same task with a weak model, needing 8,000 thinking tokens just to catch up, already costs 4 cents at the cheap tier (and yes, models have laziness too, it's just reversed in the weak one: it overthinks because it isn't sure of itself, and that costs you, not it).
For us the rule of thumb is simple: take the strongest model available and let it work at medium effort. That's almost always the sharpest and cheapest combination, simply because it needs far fewer tokens to get there.
A prompt, on the house
Before picking a model and thinking mode, check the task:
1. Deep knowledge, nuance, or judgment (writing, strategy, subtle analysis)?
→ Your strongest model, medium thinking mode.
2. A clear logical path with steps (math, debugging, planning)?
→ A mid-tier model is enough, but give it high thinking mode.
3. Both at once?
→ Your strongest model, high thinking mode. The only task worth the full price.
Not sure? Your strongest model, medium thinking mode.Next time you're torn between tier and effort, run these questions before you choose. Most of the time the answer will surprise you with how simple it is.
