StepFun AI Models: Benchmarks and Ranking

Rank	Model	Overall score	Artificial Analysis	LiveBench	Epoch AI	ARC-AGI	Scale SEAL	Vals AI	Terminal-Bench	EQ-Bench	SimpleBench	BenchLM
1	Gemini 4 Argon Google	92.4	#5	—	—	—	—	#1	—	—	—	—
2	Claude Opus 5.5 Anthropic	92.1	#1	#2	#1	#3	—	#2	—	—	#1	—
3	GPT-6 Astra OpenAI	91.6	#4	#4	#2	#1	#1	#6	#1	—	#4	#1
4	Claude Fable 5.1 Anthropic	91.1	#3	#1	#4	#6	#2	#4	#2	—	#2	#2
5	Claude Opus 5 Anthropic	89.7	#7	#10	#5	#5	—	#5	#3	#1	#9	#3
6	GPT-6.1 Sol OpenAI	89.7	#6	#5	—	#2	—	#8	—	—	#5	—
7	Claude Fable 5 Anthropic	89.6	#8	#3	#7	#8	—	#7	#4	#2	—	#4
8	Claude Sonnet 5.5 Anthropic	89.1	#2	#20	#3	—	—	#2	—	—	#14	—
9	Muse Spark 1.3 Meta	87.7	#9	#5	#17	—	—	#9	—	—	#8	—
10	GPT-5.6 Sol OpenAI	86.8	#11	#7	#8	#4	—	#10	#7	#9	#25	#5
11	GPT-6 Sol OpenAI	86.2	#10	#12	—	#7	—	#11	—	—	#17	—
12	GPT-5.5 Pro OpenAI	85.7	—	—	#6	#11	—	—	—	—	#11	—
13	MiMo V2.6 Pro Xiaomi	84.3	#13	—	—	—	—	#12	—	—	—	—
14	GPT-5.4 Pro OpenAI	84.1	—	—	#11	#15	—	—	—	—	—	—
15	Grok 4.7 xAI	83.6	#12	#21	—	—	—	#14	#6	—	—	—
16	GPT-5.5 OpenAI	83.1	#28	#9	#10	#10	—	—	—	#4	#21	#11
17	DeepSeek V4.1 Flash DeepSeek	82.2	#25	#7	#29	—	—	#22	—	—	#23	—
18	Gemini 3.8 Flash Google	82.0	#22	#32	#15	#8	#4	#15	#11	—	#6	#6
19	GLM-5.3 Z.ai (Zhipu)	81.4	#15	#30	#24	—	—	#16	#5	—	#24	—
20	Kimi K3 Moonshot AI	80.7	#18	#12	#13	#27	—	#26	—	#3	#35	#8
21	Qwen 3.8 Max Alibaba (Qwen)	80.6	#14	#15	#28	—	—	#30	—	—	—	#9
22	GPT-5.6 Terra OpenAI	80.5	#19	#19	#9	#14	—	#18	#9	#11	#55	#7
23	GPT-5.4 OpenAI	80.5	—	#16	#16	#18	—	—	—	#7	—	#13
24	Grok 4.6 xAI	80.3	#16	#16	#19	#22	—	#19	#10	—	#14	#11
25	Claude Opus 4.8 Anthropic	80.1	#20	#28	#12	#19	—	#13	#8	#6	#25	#10
26	Gemini 3.7 Flash Google	79.8	#26	#14	#14	#11	—	#22	#15	—	—	#14
27	Muse Spark 1.2 Meta	79.3	#24	#16	#30	—	—	#29	—	—	#16	—
28	MiMo V2.6 Flash Xiaomi	79.2	#31	—	—	—	—	#17	—	—	—	—
29	Step 5 StepFun	79.2	#17	—	—	—	—	#28	—	—	—	—
30	Claude Opus 4.7 Anthropic	78.3	—	#27	#22	#17	#9	—	—	#5	#28	—
31	Qwen 3.8 Flash Next Alibaba (Qwen)	76.9	#23	#28	—	—	—	—	—	—	—	—
32	Claude Sonnet 5 Anthropic	75.6	#29	#31	#23	—	—	#20	#13	#10	#36	#15
33	DeepSeek V4 Pro DeepSeek	74.0	#33	#21	#26	#26	—	#32	—	#18	#53	—
34	Grok 4.20 xAI	73.9	—	—	#42	#24	—	—	—	—	—	—
35	Gemini 3.1 Pro Google	73.3	#44	#24	#31	#16	#3	#37	—	#21	#10	—
36	Claude Opus 4.6 Anthropic	73.2	—	#38	#27	#21	#10	—	—	#13	#22	#20
37	Muse Spark 1.1 Meta	72.8	#38	#34	#35	—	—	—	—	#8	—	#16
38	Grok 4.5 xAI	72.7	#27	#32	#36	#35	—	#34	#13	—	#20	#18
39	GPT-6 Luna OpenAI	72.6	#30	#47	—	#32	—	#24	—	—	—	—
40	Muse Spark Meta	72.5	—	—	#41	—	#6	—	—	—	—	#27
41	DeepSeek V4 Flash DeepSeek	72.3	#35	#39	#32	#25	—	#31	—	—	#33	—
42	GPT-5.6 Luna OpenAI	72.1	#32	#41	#21	#31	—	#21	#12	#19	#57	#17
43	Gemini 3.5 Flash Google	71.5	#41	#36	#33	#19	—	#33	—	#24	#12	#21
44	GLM-5.3 Flash Z.ai (Zhipu)	70.6	#20	#49	#43	#23	—	—	—	—	—	—
45	Gemini 3.6 Flash Google	70.0	#36	#41	#34	#27	—	—	—	—	—	#19
46	Qwen3.8 27B Alibaba (Qwen)	68.8	#38	#34	—	#36	—	—	—	—	—	—
47	Gemini 3 Pro Google	68.0	—	—	#39	#41	#7	—	—	—	#13	#26
47	GPT-5.2 Pro OpenAI	68.0	—	—	#25	#33	—	—	—	—	#43	—
49	GPT-5.3 Codex OpenAI	67.9	—	—	#18	—	—	—	—	—	—	#24
50	GPT-5.2.11 OpenAI	67.3	—	#36	—	—	#12	—	—	—	—	—

Rank	Provider	Overall score	Top model	Models
1	Anthropic	86.3	Claude Opus 5.5	28
2	OpenAI	84.6	GPT-6 Astra	54
3	Google	79.1	Gemini 4 Argon	34
4	Meta	74.8	Muse Spark 1.3	23
5	xAI	70.8	Grok 4.7	13
6	Alibaba (Qwen)	67.2	Qwen 3.8 Max	46
7	Xiaomi	66.9	MiMo V2.6 Pro	3
8	Moonshot AI	66.6	Kimi K3	5
9	Z.ai (Zhipu)	64.7	GLM-5.3	10
10	StepFun	64.6	Step 5	1
11	DeepSeek	64.2	DeepSeek V4.1 Flash	22
12	Tencent	46.8	—	2
13	Thinking Machines	39.7	Inkling	2
14	MiniMax	37.3	Minimax M3	5
15	NVIDIA	30.5	Nemotron 3 Ultra 550B A55B	9
16	Mistral AI	18.2	Mistral 3.5	26
17	Amazon	15.9	—	3
18	Microsoft	15.7	Phi-4	6
19	Cohere	12.4	Command R	3
20	Inception	11.2	Mercury 2.5	1

A provider’s result on each site is that of its best model there.

Performance vs price

Top left = cheap and strong. Click a point for details.

Best value

Overall score ÷ price

Top open-weight models

Models with public weights

Fastest models

Output tokens per second

By country

Top 50 models

Agreement between leaderboards

Rank correlation on shared models (1 = identical).

How the ranking works

Collect — Official data from the 11 leaderboards below is fetched every hour. Scores marked as estimates are not used.
Match — Different spellings of the same model are merged. For variants such as reasoning effort, each site’s best result is used.
Standardize — Each site’s main metric is standardized within that site, keeping score gaps rather than just ranks.
Equate — A factor model, bridged by models listed on several sites, corrects for each site’s difficulty and model mix. Noisier sites get less weight.
No penalty for missing data — Models are scored only from the sites that rated them. Uncertainty is shown as a 90% interval. Models rated on 2+ sites are ranked.
Arena — Arena (formerly LMArena) is used for domain rankings only, not for the overall score. New models are added automatically.

StepFun AI models: benchmarks and ranking

More from StepFun