Skip to content

Blueprint token benchmark

On 6 small UI fixtures, measured November 20, 2026, a Figma-style JSON node tree took a median of 3.61x the tokens of the same design as Blueprint rows (range 3.42x to 4.06x), and Figma plugin JavaScript a median of 4.43x (range 4.15x to 5.21x). Tokenizer: tiktoken 0.12.0, o200k_base. The script and every fixture are on this page, so you can check the numbers.

What was measured?

The text an agent has to write to create one design, in three forms:

  • Blueprint rows. The Blueprint an agent sends to Brilliant, one element per line.

  • A Figma-style JSON node tree. Minified, holding only the fields the design needs: id, name, type, layout, size, fills, strokes, radius, effects and text style.

  • Figma plugin JavaScript. A straight-line script that builds the same tree, one statement per property, with no helpers and no error handling.

Each count is the whole text of that form, in tokens.

What are the results?

n = 6. tiktoken 0.12.0, o200k_base. Measured November 20, 2026.

FixtureElementsBlueprintFigma-style JSONvs BlueprintPlugin JavaScriptvs Blueprint
Pricing card214261,5673.68x2,0084.71x
Nav bar132459954.06x1,2765.21x
Sign-in form215051,7283.42x2,1914.34x
Dashboard tile173951,4343.63x1,7874.52x
Task list225642,0243.59x2,4274.30x
Confirm modal133821,3303.48x1,5854.15x
Median3.61x4.43x

With cl100k_base as a cross-check the medians are 3.62x and 4.46x.

What does one fixture look like?

The nav bar, 13 elements. As Blueprint, 245 tokens:

nav_bar.bl
al(h,x(sb),y(c),pad(16,32)) s(1200,hug) f[(#FFFFFF)] st[(#E5E7EB,w(1))] "Nav bar"
  al(h,y(c),g(8)) "Brand"
    c s(24,24) f[(#4F46E5)] "Mark"
    t("Acme",Inter,16,b) f[(#111827)]
  al(h,y(c),g(28)) "Links"
    t("Product",Inter,14,m) f[(#374151)]
    t("Pricing",Inter,14,m) f[(#374151)]
    t("Docs",Inter,14,m) f[(#374151)]
    t("Blog",Inter,14,m) f[(#374151)]
  al(h,y(c),g(12)) "Actions"
    t("Sign in",Inter,14,m) f[(#374151)]
    al(h,x(c),y(c),pad(8,14)) f[(#111827)] rd(8) "Button"
      t("Get started",Inter,14,sb) f[(#FFFFFF)]

The same nav bar as plugin JavaScript is 1,276 tokens. Its first 14 lines:

nav_bar.plugin.js, the first 14 lines
await figma.loadFontAsync({"family":"Inter","style":"Bold"});
await figma.loadFontAsync({"family":"Inter","style":"Medium"});
await figma.loadFontAsync({"family":"Inter","style":"Semi Bold"});
const n1=figma.createFrame();
n1.name="Nav bar";
n1.layoutMode="HORIZONTAL";
n1.primaryAxisAlignItems="SPACE_BETWEEN";
n1.counterAxisAlignItems="CENTER";
n1.paddingTop=16;
n1.paddingRight=32;
n1.paddingBottom=16;
n1.paddingLeft=32;
n1.fills=[{"type":"SOLID","color":{"r":1,"g":1,"b":1}}];
n1.strokes=[{"type":"SOLID","color":{"r":0.898,"g":0.906,"b":0.922}}];

The full files: nav_bar.figma.json and nav_bar.plugin.js.

What are the caveats?

  • We wrote the fixtures and the mapping. The six designs are generated for this benchmark, and the JSON and JavaScript forms are produced from the Blueprint by the script. Read the mapping before you trust the numbers.

  • Six fixtures is a small sample. They are a pricing card, a nav bar, a form, a dashboard tile, a list and a modal, each 13 to 22 elements. Larger or different designs can land elsewhere.

  • The Figma forms are minimal. They hold only what the design needs. A node tree exported from a real Figma file carries more fields per node, and a production plugin script carries helpers and error handling. Neither is measured here.

  • Three fixtures use design tokens and three use literal values. Both styles are in the table, so you can see each.

  • These are OpenAI tokenizers. Other models count differently, so the ratios move a little per model.

  • This measures writing a design, nothing else. It says nothing about HTML, about reading a design back, or about the quality of what an agent makes.

How do I reproduce it?

Download the script and the fixtures folder, keep them side by side, and run it with Python 3 and tiktoken:

Re-run the benchmark
pip install tiktoken
python3 bench.py

python3 bench.py --check fails if a generated fixture or results.json differs from what the script produces. Without tiktoken the script prints NOT MEASURED and no numbers.