AI judge results

Each judge sees two anonymous answers to the same native English writing prompt. It picks the stronger text—or a tie—without seeing the model names.

The tables add up every validated, published duel for each judge: wins, losses, ties and total comparisons. Blinding keeps each verdict focused on the writing rather than the brand behind it.

AI judges let us compare many model pairs with a consistent protocol. They do not replace readers: human votes are counted separately in the human ranking, so you can see where people and machines agree—or do not.

Latest included verdict: 2026-07-21 14:59 CEST

Claude Fable 5

225 validated duel verdicts

Results from Claude Fable 5
Rank	Model	Wins	Losses	Ties	Comparisons
1	Claude Fable 5 High 💡	45	0	1	46
2	MiniMax M3 💡	42	5	2	49
3	Qwen 3.7 Plus 💡	15	11	4	30
4	Qwen 3.7 Max 💡	13	11	6	30
5	MiMo V2.5 Pro 💡	13	14	4	31
6	DeepSeek V4 Pro 💡	12	12	6	30
7	Qwen 3.6 27B Q8_0 💡	9	7	2	18
8	DeepSeek V4 Flash 💡	9	19	3	31
9	MiMo V2.5 💡	7	21	3	31
10	GLM 5.2 💡	6	20	4	30
11	Gemma 4 31B Instruct QAT 💡	5	8	1	14
12	Gemma 4 12B Instruct Q8_0 (non-reasoning)	3	2	4	9
13	Qwen 3.6 27B Q4_K_M 💡	3	8	2	13
14	Qwen 3.6 27B Q8_0 (non-reasoning)	2	4	4	10
15	Gemma 4 12B Instruct Q8_0 💡	2	8	6	16
16	Qwen 3.6 35B-A3B Q8_0 💡	2	10	3	15
17	Qwen 3.6 35B-A3B Q8_0 (non-reasoning)	1	3	5	9
18	Gemma 4 12B Instruct QAT 💡	1	3	0	4
19	Gemma 4 31B Instruct Q8_0 💡	1	5	3	9
20	Qwen 3.6 35B-A3B Q4_K_M 💡	1	7	3	11
21	Gemma 4 26B-A4B Instruct Q8_0 💡	0	4	0	4
22	Gemma 4 26B-A4B Instruct QAT 💡	0	10	0	10

Claude Opus 4 8

559 validated duel verdicts

Results from Claude Opus 4 8
Rank	Model	Wins	Losses	Ties	Comparisons
1	Qwen 3.6 27B Q4_K_M 💡	71	47	9	127
2	Qwen 3.6 35B-A3B Q4_K_M 💡	71	58	5	134
3	Gemma 4 31B Instruct QAT 💡	53	69	7	129
4	MiniMax M3 💡	43	4	0	47
5	Gemma 4 26B-A4B Instruct QAT 💡	42	84	9	135
6	MiMo V2.5 Pro 💡	27	16	0	43
7	DeepSeek V4 Pro 💡	26	14	3	43
8	GLM 5.2 💡	26	15	2	43
9	Qwen 3.7 Max 💡	23	17	2	42
10	Gemma 4 26B-A4B Instruct Q8_0 💡	22	38	3	63
11	DeepSeek V4 Flash 💡	20	23	0	43
12	Gemma 4 31B Instruct Q8_0 💡	20	41	6	67
13	Qwen 3.7 Plus 💡	18	20	3	41
14	MiMo V2.5 💡	18	23	0	41
15	Qwen 3.6 27B Q8_0 💡	15	13	0	28
16	Qwen 3.6 35B-A3B Q8_0 💡	11	15	1	27
17	Gemma 4 12B Instruct Q8_0 💡	9	18	0	27
18	Gemma 4 12B Instruct QAT 💡	6	4	0	10
19	Qwen 3.6 35B-A3B Q8_0 (non-reasoning)	5	4	0	9
20	Qwen 3.6 27B Q8_0 (non-reasoning)	5	5	0	10
21	Gemma 4 12B Instruct Q8_0 (non-reasoning)	3	6	0	9

GPT-5.6 Sol

1296 validated duel verdicts

Results from GPT-5.6 Sol
Rank	Model	Wins	Losses	Ties	Comparisons
1	GLM 5.2 💡	108	47	1	156
2	DeepSeek V4 Pro 💡	97	55	3	155
3	Qwen 3.7 Plus 💡	91	59	5	155
4	Qwen 3.7 Max 💡	91	60	4	155
5	Qwen 3.6 27B Q4_K_M 💡	89	65	0	154
6	Gemma 4 26B-A4B Instruct QAT 💡	83	79	3	165
7	Qwen 3.6 35B-A3B Q4_K_M 💡	80	85	1	166
8	Gemma 4 31B Instruct QAT 💡	78	76	2	156
9	MiMo V2.5 Pro 💡	68	83	4	155
10	MiniMax M3 💡	66	29	0	95
11	DeepSeek V4 Flash 💡	63	91	1	155
12	MiMo V2.5 💡	54	101	0	155
13	Gemma 4 31B Instruct Q8_0 💡	51	42	0	93
14	Qwen 3.6 35B-A3B Q8_0 💡	47	41	0	88
15	Gemma 4 26B-A4B Instruct Q8_0 💡	42	46	1	89
16	Qwen 3.6 27B Q8_0 💡	42	46	0	88
17	Gemma 4 12B Instruct Q8_0 💡	37	51	0	88
18	Gemma 4 12B Instruct Q8_0 (non-reasoning)	37	59	2	98
19	Qwen 3.6 27B Q8_0 (non-reasoning)	21	77	0	98
20	Gpt 5.6 Sol 💡	19	3	0	22
21	Qwen 3.6 35B-A3B Q8_0 (non-reasoning)	16	82	0	98
22	Gemma 4 12B Instruct QAT 💡	2	5	1	8

The AIs have had their say. Your turn.

An AI judge can read thousands of duels without coffee or a lunch break. It still cannot tell us what makes you laugh, moves you, or keeps you reading. Your blind vote feeds a separate human ranking and helps reveal where human taste and machine judgment part ways.

Judge two anonymous texts