Significance bands

Adjacent ranks that the measurement cannot separate. The band is the object; the integer is a footnote. We do not invent a crisp #1.

93% adjacent pairs inside published error

64/134 distinct significance ranks / scored models

3 joint first — no crisp #1

LMArena publishes per-model 95% CI half-widths (CC BY 4.0). Adjacent models are inseparable when their intervals overlap.

Artificial Analysis board (evaluation only)

The live leaderboard still groups AA intelligence with a borrowed ±1.41 difference bound. On this catalogue that yields 93% of adjacent pairs (124 of 133) inside the band, and 3 joint-first models — but Artificial Analysis redistribution is unresolved, so this figure is shown for evaluation, not as a redistributable claim.

Sig. rank	Model	Elo	± half-width
1	Claude Opus 5	1504	4.8
1	Claude Opus 4.6	1503	3.5
1	Claude Fable 5	1495	5.2
3	Gemini 3.7 Flash	1490	8.1
3	Claude Opus 4.7	1490	3.9
4	Gemini 3.5 Flash	1483	4.5
4	Qwen3.8 Max	1482	6.8
5	Gemini 3.1 Pro Preview	1480	3.3
5	Muse Spark 1.1	1478	5.3
4	GLM 5.3	1477	9.9
6	Gemini 3.6 Flash	1477	5.4
5	Kimi K3	1476	5.8
9	GPT-5.5	1471	3.9
9	GPT-5.4	1470	3.9
12	GLM 5.2	1465	4.8
12	MiMo-V2.5-Pro	1465	4.0
12	GLM 5.1	1464	4.1
15	Claude Opus 4.8	1461	4.4
15	Claude Sonnet 4.6	1458	3.6
18	Gemini 2.5 Pro	1457	2.5
16	Qwen3.7 Plus	1456	4.7
18	Kimi K2.6	1455	4.5
18	GPT-5.6 Sol	1454	5.4
18	Grok 4.5	1452	5.2
21	Claude Opus 4.5	1450	3.2
20	Qwen3.6 Max Preview	1446	8.3
22	GPT-5.6 Terra	1446	5.3
24	Kimi K2.5	1445	3.3
23	GLM 5	1445	4.3
21	Grok 4.6	1444	10.1
24	Claude Sonnet 5	1442	4.9
23	Gemma 4 31B	1442	7.5
26	GPT-5.1	1441	3.7
23	Hy3	1441	8.1
22	Qwen3.8 27B	1441	10.5
26	GLM 4.6	1440	3.9
23	DeepSeek V4 Pro 0423	1439	9.8
26	Inkling	1439	5.4
27	Qwen3.5 397B A17B	1438	3.4
28	Claude Sonnet 4.5	1438	2.9
27	Qwen3.6 Plus	1437	4.2
26	GLM 5V Turbo	1437	6.7
26	Gemini 3.5 Flash Lite	1437	5.5
27	GLM 4.7	1435	6.0
29	MiniMax M3	1435	4.5
26	Gemma 4 26B A4B	1435	7.6
33	DeepSeek V4 Flash 0423	1432	4.1
35	GLM 4.5	1429	4.9