Benchmarks · composite

Arena Text · Instruction following

Crowdsourced pairwise human preference on open-ended text prompts (style control on); Bradley-Terry rating. Only the Instruction following category.

All results

216 rows

#	Model	Effort	Score
1	Gemini 4 Carbon	high	1525 ±16
2	Gemini 4 Argon	high	1525 ±10
3	Opus 5.5	high	1522 ±10
4	Opus 4.6	high	1511 ±5
5	Fable 5	high	1508 ±6
6	Opus 4.7	high	1502 ±5
7	Fable 5.1	max	1500 ±9
8	Opus 4.6	max	1498 ±5
9	Opus 5	high	1496 ±5
10	Opus 5	max	1495 ±7
11	Opus 4.7	max	1493 ±5
12	Opus 4.8	high	1492 ±5
13	Gemini 3.8 Flash	high	1491 ±6
14	GPT-5.6 Sol	xhigh	1491 ±6
15	Kimi K3	max	1490 ±6
16	Gemini 3.7 Flash	high	1487 ±7
17	Sonnet 5.5	xhigh	1486 ±11
18	GPT-6.1 Sol	max	1486 ±12
19	Opus 4.5	high	1484 ±7
20	Muse Spark 1.3	max	1483 ±8
21	Gemini 3.1 Pro	high	1480 ±4
22	GLM 5.3	max	1479 ±7
23	Opus 4.8	max	1478 ±5
24	GPT-5.5	high	1478 ±5
25	DeepSeek V4.1 Flash	max	1477 ±9
26	GPT-5.5	xhigh	1475 ±5
27	Qwen 3.8 Max	max	1475 ±7
28	Muse Spark 1.2	xhigh	1475 ±10
29	Sonnet 4.6	max	1474 ±5
30	GLM 5.3 Flash	max	1474 ±6