TinkerTwo tiers of Larkspur’s assistant: large at the back, small in front.
Chapter 4 of 7
Send it the easy ones
Larkspur mini and its student showed the mechanism. Now the real question: Larkspur’s assistant comes in two tiers, a small model and a large one, both reading the same documents. Which questions can the small tier take?
This chapter: Where does a small model answer as well as a large one, and where does it fall short?
You will learn
- How two tiers of the same assistant compare on Evals’ test set
- What a small tier costs against a large one, from a real price list
- Which kinds of question a small model gets wrong
Words you’ll meet (2)
- Tier
- One size of model offered for the same job: here, a small tier and a large tier.
- Relative cost
- Cost counted in large answers: one large answer is 1, a small one a fraction of that.
The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.
Step 1 of 3
Larkspur’s assistant comes in two tiers: a small model and a large one, both reading the same retrieved documents. A small answer costs 0.2 of a large one, from a real price list below; from here on costs are in large answers (1 each). The tiers are illustrations written for this site, not measured models. Before you run them:
Anthropic lists Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, and Claude Opus 4.5 at $5 and $25: the smaller model costs a fifth as much per token, in and out. Source: platform.claude.com (checked 2026-10-21)