glyphhunt

leaderboard

Every model, every metric, in one table. Sort any column, filter by name. Runs that left their working directory are shown but never scored.

● running

Jul 25, 2026, 07:12 AM UTC

arm64 · macOS 26.5.2 · 1920x1080 · 60fps · 1800 frames · 17 glyphs in 17 typefaces

60/72 runs 4 models ✓ 0 invalidated ~ 47 hit plan quota ✓ isolated workdirs ✓ key committed in advance

Zapfino

Papyrus

Herculanum

Chalkduster

Bodoni 72

Didot

Copperplate

Luminari

Trattatello

Snell Roundhand

Brush Script

Marker Felt

Noteworthy

American Typewriter

Impact

Phosphate

Rockwell

overall

across every model, level and prompt mode

~ partial

runs scored

exact solves

5/60

mean chars

1.77/17

mean frames

0.85/17

mean spatial

0.82/17

mean time

1726s

hit 30m wall

43/60

invalidated

every metric

click any column to sort. word, temporal and spatial accuracy are kept separate — reading the word and knowing where you saw it are different abilities.

model	solved	chars/17	accuracy	best	frame	spatial	lev	L1	L2	L3	blind	hinted	avg t	shell	ffmpeg	python	imgs	diffed	out tok	cost	fail
gpt-5.6/high	2/18	2.33	██············	17	0.56	0.56	0.18	7	0	0	3.78	0.89	1946s	46.56	10.83	21.17	0	11	20,140	—	13
opus-5	2/18	1.89	██············	17	1.89	1.89	0.11	5.67	0	0	3.78	0	1709s	47.78	11.33	44.22	24.94	6	5,310	$7.94	16
gpt-5.6/med	1/18	1.67	█·············	17	0.39	0.28	0.17	4.67	0.17	0.17	0.44	2.89	1549s	39.83	11.06	19.89	0	11	18,559	—	9
fable-5	0/6	0	··············	0	0	0	0	0	0	0	0	0	1651s	37	7	34	37.83	4	7,090	$8.34	6

what the hint is worth

blind says only that something is hidden. hinted gives the length, the one-typeface-per-letter rule, temporal ordering, and warns about faintness and decoys.

model	blind chars	hinted chars	Δ	blind solved	hinted solved
gpt-5.6/high	3.78	0.89	-2.9	2/9	0/9
opus-5	3.78	0	-3.8	2/9	0/9
gpt-5.6/med	0.44	2.89	+2.5	0/9	1/9
fable-5	0	0	0	0/3	0/3