ClutterBench AnimationsBeta GitHub

How much of the graphic is clutter?

Leaderboard

How much of each model's graphics the grader would cut. Lower is better.

Cost and time

What a graphic costs and how long it takes, against how much of it is clutter. Down and to the left is better.

Clutter against cost

Clutter against time

Pareto frontier: nothing is both cheaper and cleaneron the frontierbeatenone model across its reasoning efforts

Per graphic

Average cost

Average time

Every graphic

How it works

The task

Each model, running as an agent, reads a blog post and builds one 1200 by 1200 graphic in HTML. It browses the publisher's site for fonts, colours and logo, and may not look at the post's own figures.

Two prompts

Raw says only "make a tasteful graphic to go with the post". Simple adds plain instructions to show one idea, once, with nothing extra. The gap between them is how much a model needs to be told.

The grader

Claude Haiku 5.5 labels every element twice. Where the two passes differ, Claude Sonnet 5.5 breaks the tie. The rules were written from one person's hand labels on 24 graphics.

Code and data on GitHub


  

The rest of the prompt, the same for both