Skip to main content
Explainers · Choosing Models
A Visual Primer

There is no
“the AI”.

Families, tiers, effort levels — and why the best model for the job is often not the biggest. Six stages.

Scroll
01

Families and tiers.

Every major provider ships a family of models: a small fast one, a mid-size workhorse, and a frontier model — the biggest and most capable. Small models are startlingly good at routine work and cost a fraction of the price; frontier models earn their premium only on genuinely hard reasoning.

Small
Fast and cheap. Startlingly good at routine work — tagging, extraction, first-pass summaries — at a fraction of the cost.
Speed●●●
Cost£
Reasoning●
Mid
The workhorse. Handles most day-to-day drafting and summarising with a comfortable margin of quality.
Speed●●
Cost££
Reasoning●●
Frontier
The most capable model the provider ships. Earns its price only on genuinely hard reasoning — analysis, synthesis, judgment.
Speed●
Cost£££££
Reasoning●●●

Generic tiers, deliberately unnamed — the specific models change every few months; the shape of the family does not.

02

Capability, cost, speed.

You are always trading between three things: how clever the answer, how cheap the run, how fast the response. Frontier-quality, instant, and free is not on offer — from anyone.

CapabilityCostSpeedpick two — at best

Pick the corner that matters, accept the price at the other two.

03

The effort dial.

Modern models expose an effort setting — how hard the model thinks before it answers. Same model, different depth: low for triage, high for analysis, more still for the hardest synthesis. Effort turns one model into several tools.

low
medium
high
xhigh

The full dial, level by level, is covered in the prompting explainer — set it deliberately, not by default.

04

Match the task.

Six tasks from an ordinary week in practice. Pick one and see which tier and effort level it deserves — and why.

Choose a task ↑
05

When cheap is fine. When it never is.

✓ CHEAP IS FINE

High-volume, low-stakes, rubric-driven — and human-reviewed anyway.

  • Triage and tagging across a disclosure set
  • First-pass summaries you will read against the source
  • Internal notes and working drafts
  • Anything with a clear rubric and a checking step
✕ NEVER ECONOMISE

Where a subtly wrong paragraph costs more than every model you will ever run.

  • Advice that goes out under your name
  • Anything filed with a court or tribunal
  • Assessments of prospects and risk
  • Final citation checks before filing

The question isn’t “can the small model do it?” — it’s “what does it cost me if it gets it subtly wrong?”

06

Models change. Principles don’t.

Model names and league tables shift every few months — this page deliberately names none. What lasts:

1
Tiers exist.
Every provider ships a small, a mid and a frontier model. The names change; the shape of the family does not.
2
Effort matters.
The same model at different effort levels is effectively several different tools. Set it deliberately, not by default.
3
Match capability to consequence.
Spend on the tasks where a subtle error is expensive. Save on the tasks where it is cheap to catch.
4
Re-evaluate quarterly.
The tier boundaries move as smaller models improve. What needed the frontier last year may not this year.
A practical suggestion: keep a one-page internal note of which tool your practice uses for which class of task — and review it each quarter.

The right tool, every time.

One question remains in the series: where did all this capability come from? Next: how models are trained — and why they behave the way they do.

Next: How Models Are Trained →