50 models · one request · in isolation

Random
4×4

Asked for a random 4×4 matrix, the models don’t spin a wide net of numbers. They reach into the same small drawer — single-digit integers — over and over.

PROMPT“Give me a random 4x4 matrix of numbers.”
41matrices parsed
45%entries are single digits 1–9
57%entries fall in 0–12
0two models gave the same matrix
01

Where the numbers land

all 656 entries

A genuinely random matrix over, say, 1–100 would spread evenly. These don’t. 296 of 656 entries are single digits 1–9, and the mass piles up in 0–12 with a long, thin tail out to 100 (and 104 stray negatives, 31 zeros). Black bars are the 0–9 band.

35<8-514-418-316-213-13203313823732743253363173283491010151121124131114715216517218819420134>
Frequency of each integer value across every matrix. Tails (<−5, >20) are bucketed at the ends.
02

Every matrix

41 specimens · cell darkness = magnitude

No two models produced the same matrix — yet they all look alike: pale grids of little numbers. Each cell is shaded by magnitude, so large values show up as dark squares. There aren’t many.

1100478312653915824763419504268
deepseek-v4-flash
512-381-741593-162107-4
mimo-v2.5
7-3120-582-146-9311-251
minimax-m3
12478935612217893354271194066
hy3-preview
7-312529-18-4611-2101-74
claude-opus-4.7
7294183650273918
claude-opus-4.8
14-3821512-9027611-421879
glm-5.2
73-512-291614-84035-112
claude-sonnet-4.6
7-214059-631218-4-96211
gpt-5.5
110023871254641983375196284457
gemini-3-flash-preview
01073042915684215
deepseek-v3.2
143-79-280115-461122-53
glm-5.1
7-251119-30-638-410-726
gemini-2.5-flash-lite
9321751456882371123942860395
gemini-2.5-flash
-372-540-816-293-75-14
nemotron-3-ultra-550b-a55b
505012-3744-9-2123-15318-439-28-4519
gpt-oss-120b
3-15204-27-681943-70
laguna-m.1
42176385531974264836011893752
mimo-v2.5-pro
4-13205-21630-42-378
gpt-4o-mini
12014319871211552094116
gemini-3.1-flash-lite
7-214509-311618-413-7210
gpt-5.4
7143221851191622741021613
nex-n2-pro
-1010-37052-84-161-59-23
nemotron-3-super-120b-a12b
7312591418210611134150
claude-opus-4.6
143278511942611352482293115
gemma-4-26b-a4b-it
3.2-1.58.70.46.14.3-2.85.9-0.79.23.6-4.17.4-3.21.82.5
claude-haiku-4.5
11005387299234657841175632897123
mistral-nemo
7-25104-3869-12-5370
gpt-5.4-mini
110023871254619437638925217048
gemma-4-31b-it
2-570-38-1460-92-435-7
kimi-k2.7-code
120143197211516820112159
gemini-3.1-flash-lite-preview
7-212049-53-168-32111-4
glm-5
42178859363174295512816839947
qwen3.7-plus
-10107-305-284-91-637-82
kimi-k2.5
0938160472953862
gpt-oss-20b
148329576194472389512650278831
qwen3.7-max
3-17052-49-681-34062
gpt-5.4-nano
7-312592-8154116-1-61308
claude-sonnet-4.5
3782156491437629583884722715419
laguna-xs.2
0952819407362581
grok-4.3
3-712059-28-461-1110-372
gpt-5.2
Hover a bar in figure 01 to mark every cell holding that value ↑
03

Digit 0–9 distribution

within the small-integer drawer

Inside that 0–9 band the spread is fairly even — so it isn’t a single favourite digit (as with the lone “pick a number” test). The bias here is magnitude: a “random” matrix that almost never leaves single figures.

031
133
236
336
425
532
631
730
832
932
04

References