Tier 2 · Practitioner2.515 min

Which model — and when the top tier earns its keep

A grassy walking track forking through open country

Every so often a more powerful model appears, and you have to decide whether your work justifies it — the top tier, or whether a cheaper one will do the job just as well. Most people decide this in one of two ways, and both are poor.

The first is “always use the best.” Comfortable, expensive, and it teaches you nothing about where the money is actually doing work. The second is “vibes” — the task feels hard, so it gets the big model. The trouble is that how hard a task feels correlates badly with the specific kind of difficulty a stronger model actually fixes. Most tasks aren’t limited by the model’s capability at all; they’re limited by a vague brief or messy inputs — and a bigger model can’t fix either of those.

There’s a better way, and it’s the same discipline this whole course teaches, pointed at a new decision: score the task on the characteristics that genuinely reward a stronger model, and pay for the top tier only where those characteristics are present.

Where a stronger model earns its keep

A small set of task features reliably reward capability. Score each separately — a blended “how much would the big model help?” just invites vibes back in.

Reasoning leverage. Long chains where an early mistake quietly poisons everything downstream; problems holding many interacting constraints at once. If an error in step 2 corrupts step 14, tier matters.
Synthesis and research leverage. Reconciling sources that disagree; adversarial reading — finding what a document leaves out, not just summarising what it says. Weaker models summarise; stronger ones notice the missing piece of the argument.
Strategic depth. Work where a mediocre answer isn’t wrong, just shallow — an analysis that’s accurate but misses the framing that would have changed your decision. Shallowness passes a quick check and costs you in the real world, which is exactly why it’s worth paying to avoid.

And two things that are not about the model tier, though people wrongly reach for a bigger model to fix them:

Making things up. Score how costly a fabricated claim would be — but don’t treat the top tier as the cure. Hallucination is held down by grounding: giving the model sources, requiring citations, and keeping your own eyes on the result (Lesson 1.4). A framework that whispers “the big model won’t invent things” is teaching the wrong habit.
Bias. A stronger model does not reliably produce fairer output. If a task’s main risk is biased output, the fixes are in the process — diverse sources, an explicit counter-argument, review by people who’d see it differently — not in the price of the model.

The axis about your material, not the model

One more axis, and it scores your inputs rather than the task: readiness. How well-specified is the job, and how clean is what you can hand the model? A brilliantly capability-sensitive task with a vague brief and contradictory source documents produces expensive confusion, not brilliance. Readiness tells you whether the task is ready for any model — and, among equally strong candidates, do the ones with clean inputs first.

Then simple economics: a one-off report barely cares about tier pricing; a task you run every day compounds it.

The rule that keeps it honest

This is the course’s spine, applied to a new decision: every judgment cites something. Where you can’t see enough to score a task, the right output is a flag — “not enough information; here’s what I’d need” — never a confident guess dressed as a verdict. The failure mode of any triage is a fluent table of scores resting on nothing. You already demand evidence from Claude; demand it of the decision about which Claude to use.

Naming the tier — carefully

At the time of writing, Anthropic’s most capable widely-released model is Claude Fable 5, sitting above the Opus, Sonnet, and Haiku tiers. But that’s exactly the kind of fact that dates: model names, capabilities, and pricing change often. The one claim worth building on is the durable one — a higher tier is more capable than the ones below it. For anything more specific — what’s current, what it costs — check Anthropic’s own pages (anthropic.com/news, docs.claude.com) rather than trusting a model, or this page, from memory. Refusing to build a decision on an unverified claim is the first lesson of the whole course, not a footnote to this one.

The move

A task is a candidate for the top tier when it scores high on at least two of reasoning, synthesis, and strategic depth. High hallucination-exposure alone isn’t enough — that’s a grounding job, not a tier job. Say “not worth it” plainly where it’s true; most work doesn’t need the top tier.

And the triage is a hypothesis, not a verdict. For your top candidate, the honest next step is cheap: run the task once on your current tier and once on the top tier, and compare the two outputs yourself. The triage tells you where that experiment is worth the cost; the experiment tells you the truth.

Think time

Think of the most demanding thing you asked Claude to do this month. Score it on reasoning, synthesis, and strategic depth. Did it clear the bar for the top tier — or was it actually a readiness problem, a vague brief a bigger model would only have failed more expensively?