experiment / language models in pixels

one prompt and a 32×32 canvas for comparing how language models think in pixels

price range
time range
Result

Loading results...

user prompt

Loading the user prompt...

system instruction

Loading the system instruction...

structured output

Strict JSON Schema (pixel_image) enforced by OpenRouter.

method

Loading the exhibition method...

setup

Zero-shot and single-turn. Models received no example image, example answer, or conversation history. A local Node.js harness sent the same prompt, recorded time and OpenRouter cost, and validated the 32×32 result.

measurement

Price is the OpenRouter-reported API cost. Time is elapsed request time.

model intelligence

Artificial Analysis Intelligence Index, snapshot 2026-08-22. Each score matches the exact model and thinking level tested by Artificial Analysis. No score is shown when that combination was not tested.

output limit

Output limit reached means the model or provider stopped the response before it completed all 32 pixel rows, so there was no valid image to display.

date

total reported cost

Loading...

runner

download the benchmark runner

Unofficial fan experiment. Super Mario is a Nintendo character. Pixel32Bench is not affiliated with Nintendo.

see Pixel16Bench
back to experiments