Which model tier runs your agent

Agents at Work — CC BY 4.0

The last lesson asked one question about your model — whose computer is it running on? (custody). This one asks the other half: which tier of model should run this agent? When you’re running one agent it barely matters. When you’re running several, or one of them daily, it becomes a real decision — about cost and about quality — and it’s Anchor 2, continuous improvement, in a very concrete form: put the capability where it earns its keep, and not a dollar more.

The two instincts that both fail

“Always use the best.” Comfortable, expensive, and it teaches you nothing about where the money is doing work. “Vibes” — this agent feels important, so it gets the top model. But how important an agent feels correlates badly with the specific kind of difficulty a stronger model actually addresses. Most agents aren’t limited by the model’s capability; they’re limited by a vague job or messy inputs, and a bigger model fixes neither.

The disciplined answer is the same triage the whole course has taught, pointed at your gallery: score each agent on the characteristics that genuinely reward a stronger model, and pay the top tier only where they’re present.

Where the top tier earns its keep — across your gallery

• Reasoning leverage — long chains where an early error compounds silently. A Bookkeeper reconciliation-checker running one written rule is low leverage; a cheaper tier does it well. An agent making a call that chains across many documents is high.
• Synthesis leverage — reconciling sources that disagree, and adversarial reading (what a source leaves out). Your Market or Competitive Analyst, weighing conflicting reports, is where the top tier shows.
• Strategic depth — where a mediocre answer isn’t wrong, just shallow, and the shallowness costs you. A cheaper model summarises; a stronger one notices the framing that changes your decision.

And the two things a bigger model will not fix — which the earlier tiers already taught you:

• Making things up is held down by grounding — sources, criteria, a human check (Tier 3’s guardrails), not by the price of the model.
• Bias does not reliably shrink with tier. The Recruiter’s lesson holds: you don’t buy your way out of a fairness problem with a more expensive model — you scope it, test it, and sometimes decline it.

Then two plain checks: readiness — a capability-hungry agent handed a vague job produces expensive confusion, not brilliance — and volume — tier pricing barely matters for a once-a-week agent and compounds for one that runs all day.

The move that keeps it model-agnostic

Here’s what makes this sit cleanly beside the last lesson rather than fighting it: the framework doesn’t care whose model it is. It tells you where capability earns its keep — and that’s just as true of the sovereign, New-Zealand- or EU-hosted models from Lesson 3.4 as of any public frontier tier. So the two questions compose into one grid:

• Whose computer (custody) — decided by what the agent touches.
• Which tier (capability) — decided by whether the agent’s work rewards a stronger model.

A sensitive-data agent belongs on sovereign infrastructure whatever its tier; a capability-hungry agent on non-sensitive work can reach for the strongest tier available. You allocate on both axes, deliberately, rather than defaulting the whole fleet to the most expensive box.

Naming the tier — carefully

At the time of writing, the most capable widely-released model is Claude Fable 5, above the Opus, Sonnet, and Haiku tiers — but that is exactly the kind of fact that dates: names, capabilities, and prices change often, and the sovereign options move too. The durable claim is simply that a higher tier is more capable than the ones below it. For current specifics, check the source (anthropic.com/news, docs.claude.com) rather than trusting a course page from memory — the same evidence discipline you’d demand of the agent itself. (This course’s legislative-watch keeps an eye on when these facts shift.)

The build move

• Triage each agent on reasoning, synthesis, and strategic depth. Two of three high → a candidate for the top tier. Otherwise a cheaper tier, honestly.
• Don’t let hallucination-exposure or bias push the tier up — those are grounding and scope jobs.
• Then run the cheap experiment: for a candidate, run it once on your current tier and once on the top, and compare the outputs yourself. The triage says where the experiment is worth it; the experiment tells the truth.

Next

That’s the guard-and-build tier done: scope, criteria, guardrails, testing, the two builds, and the two questions about your model — whose computer, and which tier. Tier 4 puts the agent to work and keeps you answerable for it.