Two variables, one dot each.
import { Scatter } from "diagrammatic";Twelve models. X is dollars per million output tokens. Y is the FrontierBench 1.1 score. Area is context window in thousands. The third channel is why this is not a line.
- glacier · 24.5 · $0.42 · 128k
- pico · 27.9 · $0.61 · 32k
- quill · 31.6 · $1.1 · 128k
- atlas-0 · 40.2 · $3.2 · 200k
- swift · 45.8 · $1.9 · 128k
- nova · 46.3 · $4.1 · 200k
- atlas-1 · 49.1 · $9.6 · 1000k
- atlas-1.1 · 51.2 · $12.4 · 1000k
- tiny · 18.4 · $0.28 · 8k
- nova-fast · 44.1 · $6.8 · 256k
- kite · 38.6 · $2.4 · 64k
- ridge · 47 · $8.2 · 400k
Illustrative data. FrontierBench 1.1, list prices per million output tokens.
swift sits at 45.8 for $1.90 with 128k context. atlas-1 scores 3.3 points higher at roughly five times the price and a 1M window. No plotted model costs under $2 while scoring above 50.
A volcano. X is log2 fold change. Y is −log10 p. The guides mark the fold-change and p-value cuts.
Illustrative data. Differential-expression results, treated versus vehicle.
Six points clear both cuts on the right, while five clear them on the left. The highest point sits on the right at a log2 fold change of 2.6 and −log10 p of 11.2.
- Use it for
- Relationships, clusters, and outliers between two measures.
- Watch out
- The trend line is a summary, not proof; correlation is not causation.