fontbench tests whether vision models can identify a typeface from a rendered image. Each item shows one specimen and four candidate font families; the model picks one.
Results are from 60 items (three seeded runs of 20) on the private mixed dataset. Runs use the Cursor SDK, so token counts are a cost proxy rather than provider API prices. See the live benchmark for methodology and context.
Requires Bun.
bun install
bun run dataset:all
bun run e2b:eval -- --model gemini-3.1-pro --limit 20See DATASET.md for dataset and evaluation details.