L-systems are fun ways to get very varied output

In short

Models drawing a half pruned coffee tree

L-system: hedge-pruned coffee shrub, by deepseek-v4.1-flash
DeepSeek 4.1 Flash, run 1. Every left branch is a short stub.$0.27SVGshrub.json
L-system: hedge-pruned coffee shrub, by gpt-6-luna
GPT-6 Luna, run 2. Nothing grows right of the trunk.$0.01SVGshrub.json
L-system: hedge-pruned coffee shrub, by gemini-3.8-flash
Gemini 3.8 Flash, run 1. Is this a tree?$0.45SVGshrub.json

Finding

L-systems are a fun way to see how models go about doing things

The prompt, v1, frozen

Write an L-system as JSON with the keys axiom, rules, angle, iterations and start_heading for a coffee shrub that has been hedge-pruned flat along one side. F and G draw forward, f moves without drawing, + and - turn by the angle, [ and ] save and restore position.

What I checked

The checker fails two fakes: half a tree, or just a line on the side to emulate a shrubed tree.

DeepSeek drew a tree with clear prunned branches on the left and non-pruned on the right. Luna also drew nice trees. Qwen3.8 Flash did not draw much as it kept measuring: in two runs it wrote its own scripts to score rules for flatness and hit the twelve minutes before rendering an answer. Gemini's three runs broke on provider errors, and mostly produced not so great looking trees.

I added the half-shrub test after these runs.

The 10 runs cost $2.03.

Older, 27 Sept

Does measuring help a model lay out a Venn? With a bit more statistics.

Newer, 29 Sept

Sonnet 5.5 against Opus 5.5, which one costs more and perform bests on the Venn, the French press and the shrub?