Model Selection: Fast vs Deep, and When Each Wins
The question is never which model is smarter — it's how much thinking a specific task actually needs, and routing before you type beats deciding after you're disappointed.
Anthropic doesn't ship one Claude — it ships several at once, trading speed and cost for depth of reasoning. Most people either ignore this entirely and let the default pick for them, or overcorrect and reach for the most powerful option for everything, treating "deepest thinking available" as a synonym for "best." Both habits leave real time and quality on the table.
The Ladder, Not the Leaderboard
The mistake baked into "which model is smartest" is that it frames tier selection as a competition with one winner. It isn't. Each tier is genuinely better suited to a different shape of task, and the right question is never "which is smarter" — it's "how much thinking does this specific task actually need."
The fast tier isn't a weaker version of the others — it's the correct tool for high-volume, well-defined work: classifying, reformatting, quick extraction, anything where the task is unambiguous and speed matters more than depth. The balanced tier is the right default for the bulk of everyday work — writing, research, and analysis that don't require unusual depth, which describes most of what you'll actually ask for in a normal week. The deep tier earns its cost on long, ambiguous, high-stakes reasoning — the kind of task where a sharp colleague would visibly pause before answering, because getting it wrong actually matters.
The specific names attached to each tier will keep changing — that's true of every AI lab, not a Claude-specific quirk. What doesn't change nearly as often is the shape: a fast tier, a balanced default, and a deep-reasoning tier, in that order. Learn the shape and you'll know how to route the moment a new lineup ships, without having to relearn the whole framework from scratch.
Routing Before You Type, Not After You're Disappointed
The habit worth building isn't "try the default, and switch tiers if it's bad." That's reactive — you pay for a wasted first attempt every time you guess wrong. The better habit is routing the task before you write the prompt, using one quick test.
The one-question test is deliberately simple: could a sharp colleague answer this correctly on the first try, without pausing to think? If yes, that's a fast-tier task — the mechanical reformatting, the one-line summary, the task where thinking harder doesn't actually change the answer. If they'd want a real minute to think it through — untangling a genuinely ambiguous tradeoff, reviewing a long document for a risk that isn't obvious on the surface — that's your deep-tier signal, and reaching for the balanced default here just means you'll likely redo the request anyway once the shallow version disappoints you.
Most tasks, honestly, land in the balanced middle, which is exactly why it's the default and not an edge case. The routing decision that actually matters day to day isn't balanced-versus-deep — it's noticing the small slice of fast-tier-shaped work you've been unconsciously over-thinking by using the default tier out of habit, and the smaller slice of genuinely hard work you've been under-thinking the same way.
Escalating Mid-Task Costs Nothing
One thing worth being explicit about: switching tiers mid-conversation is free in every way that matters. If you started at the balanced default and the output feels shallow — vague where it should be specific, missing an obvious tradeoff — escalating to the deep tier for the next message costs you nothing but the time to notice it. There's no penalty for guessing wrong once and correcting course.
That's exactly why "start at the default, escalate on disappointment" is a perfectly good fallback for the genuinely ambiguous cases where the one-question test doesn't give you a clean answer. The real cost isn't escalating late — it's the opposite pattern, where indecision about which tier to use burns more time than either tier would have cost on its own. When you're stuck deciding, that hesitation is itself the signal to just start at the balanced tier and let the first response tell you whether you need to go deeper.
With Projects, artifacts, Styles, and now model routing all in place, the next few lessons turn from how the tool works to what you actually do with it — starting with research, where the model-tier question gets a real answer: verification work is where the deep tier tends to earn its keep.