战绩榜 MODEL ARENA

评测讲「考什么」，战绩榜讲「谁考了多少分」——当前旗舰模型的实测成绩

评测卡目录 · 基准档案战绩榜 · 36 个模型

数据源：Artificial Analysis（统一协议复测）抓取于 2026-10-19

36 个旗舰模型 18 个 benchmark 维度 8 个已挂评测卡 18 家厂商 · 16 个开源权重

当前综合指数领跑

Claude Opus 5.5

Anthropic · 闭源 · Max 档

57.6

AA 综合智能指数

综合指数排行

同一打分体系下的可比分数；所有 benchmark 均由 Artificial Analysis 统一复测，不是厂商自报。

Claude Opus 5.5

Anthropic · Max

57.6

Claude Sonnet 5.5

Anthropic · Max

56.0

Muse Wonder Full

Meta · Max

55.8

Claude Fable 5.1

Anthropic · Max

53.4

GPT-6 Astra

OpenAI · Max

52.7

Gemini 4 Argon

Google · High

52.6

Muse Wonder

Meta · Max

52.4

GPT-6.1 Sol

OpenAI · Max

51.8

Muse Spark 1.3

Meta · Max

48.1

Grok 4.7

xAI · Xhigh

46.5

MiMo-V2.6-Pro

小米开源

46.3

Qwen3.8 Max

阿里通义 · 0902

45.4

GLM-5.3

智谱 · Max 开源

44.8

Step 5 Preview

阶跃星辰开源

43.7

Kimi K3

月之暗面 · Max 开源

43.6

Claude Haiku 5.5

Anthropic · Max

43.4

GPT-5.6 Terra

OpenAI · Max

42.1

GLM-5.3-Flash

智谱开源

41.8

Ling 3.1 Flash

InclusionAI 开源

41.1

Gemini 3.8 Flash

Google · High

40.9

Qwen3.8 2.4T A95B

阿里通义开源

40.0

Qwen3.8-Flash-Next

阿里通义开源

39.8

DeepSeek V4.1 Flash

DeepSeek · Max 开源

39.5

Muse Spark Open

Meta · Xhigh 开源

38.9

Mistral Large 4 Preview

Mistral

38.4

GPT-6 Luna

OpenAI · Max

38.1

MiMo-V2.6-Flash

小米开源

37.9

DeepSeek V4 Pro 0813

DeepSeek · Max 开源

36.0

Agnes 2.5 Pro Beta

Sapiens AI

35.2

DeepSeek V4 Flash Vision

DeepSeek · Max

34.8

JT-4.1 Flash 236B A21B

China Mobile · Reasoning

33.9

Qwen3.8 27B

阿里通义 · Xhigh 开源

33.7

Motif 3

Motif Technologies 开源

33.6

GPT-5.3 Codex

OpenAI · Xhigh

32.5

K2 Horizon 375B A23B

Institute of Foundation Models 开源

30.5

Beam 1.0

Reflection 开源

30.3

Benchmark 维度矩阵

点列头可按该维度排序；列名可点进对应的评测卡。空格代表该维度尚未对该模型复测。

AA 用统一协议合成的可比分数

模型	AA Intelligence Index 综合智能指数	AA Omniscience 全知指数
Claude Opus 5.5 Anthropic · Max	57.6	+46.4
Claude Sonnet 5.5 Anthropic · Max	56.0	+32.3
Muse Wonder Full Meta · Max	55.8	+30.1
Claude Fable 5.1 Anthropic · Max	53.4	+43.5
GPT-6 Astra OpenAI · Max	52.7	+43.4
Gemini 4 Argon Google · High	52.6	+42.4
Muse Wonder Meta · Max	52.4	+25.5
GPT-6.1 Sol OpenAI · Max	51.8	+41.5
Muse Spark 1.3 Meta · Max	48.1	+25.0
Grok 4.7 xAI · Xhigh	46.4	+32.0
MiMo-V2.6-Pro 小米	46.3	+8.4
Qwen3.8 Max 阿里通义 · 0902	45.4	+12.0
GLM-5.3 智谱 · Max	44.8	+14.3
Step 5 Preview 阶跃星辰	43.7	+16.4
Kimi K3 月之暗面 · Max	43.6	+19.7
Claude Haiku 5.5 Anthropic · Max	43.4	+10.7
GPT-5.6 Terra OpenAI · Max	42.1	+0.1
GLM-5.3-Flash 智谱	41.8	+7.5
Ling 3.1 Flash InclusionAI	41.1	+2.2
Gemini 3.8 Flash Google · High	40.9	+29.6
Qwen3.8 2.4T A95B 阿里通义	39.9	+4.3
Qwen3.8-Flash-Next 阿里通义	39.8	-9.7
DeepSeek V4.1 Flash DeepSeek · Max	39.5	-5.3
Muse Spark Open Meta · Xhigh	38.9	+24.6
Mistral Large 4 Preview Mistral	38.4	-5.3
GPT-6 Luna OpenAI · Max	38.1	+0.7
MiMo-V2.6-Flash 小米	37.9	-12.7
DeepSeek V4 Pro 0813 DeepSeek · Max	36.0	+0.8
Agnes 2.5 Pro Beta Sapiens AI	35.2	-10.5
DeepSeek V4 Flash Vision DeepSeek · Max	34.8	-17.6
JT-4.1 Flash 236B A21B China Mobile · Reasoning	33.9	+20.6
Qwen3.8 27B 阿里通义 · Xhigh	33.7	-10.0
Motif 3 Motif Technologies	33.6	+10.2
GPT-5.3 Codex OpenAI · Xhigh	32.5	+10.9
K2 Horizon 375B A23B Institute of Foundation Models	30.5	-3.0
Beam 1.0 Reflection	30.3	+1.5

通用知识、科学、多模态理解与科研编程

模型	GPQA Diamond 研究生级科学问答	Humanity's Last Exam 终极考试	MMMU-Pro 多模态理解 Pro	SciCode 科研编程
Claude Opus 5.5 Anthropic · Max	—	61.4%	87.7%	66.9%
Claude Sonnet 5.5 Anthropic · Max	—	55.0%	—	61.0%
Muse Wonder Full Meta · Max	—	53.6%	—	62.6%
Claude Fable 5.1 Anthropic · Max	93.7%	59.1%	—	63.1%
GPT-6 Astra OpenAI · Max	96.1%	54.7%	86.9%	56.5%
Gemini 4 Argon Google · High	—	57.1%	—	61.8%
Muse Wonder Meta · Max	—	50.1%	—	60.3%
GPT-6.1 Sol OpenAI · Max	—	52.9%	86.0%	54.2%
Muse Spark 1.3 Meta · Max	93.5%	48.7%	—	58.8%
Grok 4.7 xAI · Xhigh	—	43.1%	—	57.4%
MiMo-V2.6-Pro 小米	—	49.4%	—	60.9%
Qwen3.8 Max 阿里通义 · 0902	92.8%	43.1%	82.8%	52.1%
GLM-5.3 智谱 · Max	91.7%	42.3%	—	59.0%
Step 5 Preview 阶跃星辰	—	46.5%	76.4%	58.9%
Kimi K3 月之暗面 · Max	93.5%	46.9%	80.5%	59.5%
Claude Haiku 5.5 Anthropic · Max	—	44.4%	—	55.0%
GPT-5.6 Terra OpenAI · Max	92.5%	42.9%	80.7%	55.0%
GLM-5.3-Flash 智谱	91.2%	39.9%	—	51.6%
Ling 3.1 Flash InclusionAI	—	39.4%	—	54.0%
Gemini 3.8 Flash Google · High	95.3%	47.8%	85.6%	56.6%
Qwen3.8 2.4T A95B 阿里通义	93.5%	42.4%	—	54.0%
Qwen3.8-Flash-Next 阿里通义	92.3%	38.0%	79.8%	50.6%
DeepSeek V4.1 Flash DeepSeek · Max	—	39.3%	77.0%	51.8%
Muse Spark Open Meta · Xhigh	89.0%	43.9%	80.9%	58.5%
Mistral Large 4 Preview Mistral	—	35.0%	76.4%	54.2%
GPT-6 Luna OpenAI · Max	—	38.5%	79.7%	54.6%
MiMo-V2.6-Flash 小米	—	35.1%	73.1%	51.3%
DeepSeek V4 Pro 0813 DeepSeek · Max	92.8%	41.0%	—	51.0%
Agnes 2.5 Pro Beta Sapiens AI	90.5%	37.5%	75.4%	48.8%
DeepSeek V4 Flash Vision DeepSeek · Max	91.3%	34.5%	74.8%	49.6%
JT-4.1 Flash 236B A21B China Mobile · Reasoning	—	43.6%	—	52.0%
Qwen3.8 27B 阿里通义 · Xhigh	90.5%	33.9%	76.3%	46.6%
Motif 3 Motif Technologies	83.4%	37.0%	—	42.3%
GPT-5.3 Codex OpenAI · Xhigh	91.5%	42.5%	78.5%	—
K2 Horizon 375B A23B Institute of Foundation Models	87.3%	32.0%	—	42.9%
Beam 1.0 Reflection	—	34.7%	—	48.5%

可验证约束下的按要求执行

模型	IFBench 指令跟随
Claude Opus 5.5 Anthropic · Max	—
Claude Sonnet 5.5 Anthropic · Max	—
Muse Wonder Full Meta · Max	—
Claude Fable 5.1 Anthropic · Max	—
GPT-6 Astra OpenAI · Max	—
Gemini 4 Argon Google · High	—
Muse Wonder Meta · Max	—
GPT-6.1 Sol OpenAI · Max	—
Muse Spark 1.3 Meta · Max	—
Grok 4.7 xAI · Xhigh	—
MiMo-V2.6-Pro 小米	—
Qwen3.8 Max 阿里通义 · 0902	—
GLM-5.3 智谱 · Max	—
Step 5 Preview 阶跃星辰	—
Kimi K3 月之暗面 · Max	—
Claude Haiku 5.5 Anthropic · Max	—
GPT-5.6 Terra OpenAI · Max	71.2%
GLM-5.3-Flash 智谱	—
Ling 3.1 Flash InclusionAI	—
Gemini 3.8 Flash Google · High	—
Qwen3.8 2.4T A95B 阿里通义	—
Qwen3.8-Flash-Next 阿里通义	—
DeepSeek V4.1 Flash DeepSeek · Max	—
Muse Spark Open Meta · Xhigh	—
Mistral Large 4 Preview Mistral	—
GPT-6 Luna OpenAI · Max	—
MiMo-V2.6-Flash 小米	—
DeepSeek V4 Pro 0813 DeepSeek · Max	—
Agnes 2.5 Pro Beta Sapiens AI	—
DeepSeek V4 Flash Vision DeepSeek · Max	—
JT-4.1 Flash 236B A21B China Mobile · Reasoning	—
Qwen3.8 27B 阿里通义 · Xhigh	—
Motif 3 Motif Technologies	—
GPT-5.3 Codex OpenAI · Xhigh	75.4%
K2 Horizon 375B A23B Institute of Foundation Models	—
Beam 1.0 Reflection	—

工具调用、终端操作、企业流程与运维

模型	τ²-bench 工具调用智能体	τ²-bench (Banking) 金融工具智能体	Terminal-Bench (Hard) 终端智能体·困难集	Terminal-Bench 2.0 终端智能体 2.0	Terminal-Bench 4.0 终端智能体 4.0	Apex-Agents 企业流程智能体	ITBench (SRE) 运维智能体	Analyst Agent 分析智能体	GDPval 经济价值任务
Claude Opus 5.5 Anthropic · Max	—	—	—	—	59.6%	—	38.2%	56.3%	68.3%
Claude Sonnet 5.5 Anthropic · Max	—	—	—	—	63.6%	—	—	57.5%	66.9%
Muse Wonder Full Meta · Max	—	—	—	—	57.1%	—	—	—	67.8%
Claude Fable 5.1 Anthropic · Max	—	47.2%	—	91.4%	52.0%	—	49.5%	57.5%	62.7%
GPT-6 Astra OpenAI · Max	—	41.4%	—	88.4%	59.1%	—	48.6%	51.2%	53.6%
Gemini 4 Argon Google · High	—	—	—	—	57.1%	—	—	48.8%	56.2%
Muse Wonder Meta · Max	—	—	—	—	52.0%	—	—	—	63.5%
GPT-6.1 Sol OpenAI · Max	—	—	—	—	56.1%	—	—	50.0%	54.7%
Muse Spark 1.3 Meta · Max	—	50.5%	—	84.3%	33.3%	—	33.2%	37.5%	59.1%
Grok 4.7 xAI · Xhigh	—	—	—	—	25.8%	—	42.1%	52.5%	60.6%
MiMo-V2.6-Pro 小米	—	—	—	—	34.8%	—	—	37.5%	59.2%
Qwen3.8 Max 阿里通义 · 0902	—	47.8%	—	88.8%	38.9%	42.4%	40.3%	45.0%	58.3%
GLM-5.3 智谱 · Max	—	50.3%	—	83.9%	41.9%	—	46.1%	38.8%	57.5%
Step 5 Preview 阶跃星辰	—	—	—	—	33.3%	38.0%	55.6%	35.0%	54.3%
Kimi K3 月之暗面 · Max	—	46.0%	—	85.0%	12.6%	41.3%	47.7%	38.8%	51.6%
Claude Haiku 5.5 Anthropic · Max	—	—	—	—	32.8%	—	—	53.8%	55.9%
GPT-5.6 Terra OpenAI · Max	86.3%	40.2%	57.6%	88.0%	35.4%	38.9%	51.0%	—	47.6%
GLM-5.3-Flash 智谱	—	47.2%	—	84.3%	32.8%	—	51.2%	25.0%	57.1%
Ling 3.1 Flash InclusionAI	—	—	—	—	33.3%	—	—	—	55.9%
Gemini 3.8 Flash Google · High	—	45.0%	—	87.6%	19.7%	—	52.5%	46.3%	46.8%
Qwen3.8 2.4T A95B 阿里通义	—	49.1%	—	82.0%	11.1%	—	—	—	55.4%
Qwen3.8-Flash-Next 阿里通义	—	45.4%	—	86.1%	25.3%	—	—	—	56.4%
DeepSeek V4.1 Flash DeepSeek · Max	—	—	—	—	26.8%	—	46.9%	35.0%	55.0%
Muse Spark Open Meta · Xhigh	—	—	—	76.4%	8.1%	—	—	—	45.1%
Mistral Large 4 Preview Mistral	—	—	—	—	26.8%	—	—	—	46.1%
GPT-6 Luna OpenAI · Max	—	—	—	—	12.6%	—	—	43.8%	46.6%
MiMo-V2.6-Flash 小米	—	—	—	—	22.7%	—	—	—	55.4%
DeepSeek V4 Pro 0813 DeepSeek · Max	—	39.6%	—	78.6%	14.1%	—	—	—	47.5%
Agnes 2.5 Pro Beta Sapiens AI	—	35.7%	—	69.7%	—	—	—	—	41.4%
DeepSeek V4 Flash Vision DeepSeek · Max	—	41.0%	—	74.2%	12.1%	—	—	—	52.3%
JT-4.1 Flash 236B A21B China Mobile · Reasoning	—	—	—	—	3.0%	—	—	—	41.2%
Qwen3.8 27B 阿里通义 · Xhigh	—	48.0%	—	79.8%	5.6%	—	—	—	46.0%
Motif 3 Motif Technologies	—	35.3%	—	74.9%	—	—	—	—	30.9%
GPT-5.3 Codex OpenAI · Xhigh	86.0%	—	53.0%	—	—	—	—	—	—
K2 Horizon 375B A23B Institute of Foundation Models	—	34.2%	—	71.9%	1.5%	—	—	—	43.0%
Beam 1.0 Reflection	—	—	—	—	3.0%	—	—	—	33.9%

长文本跨段检索与推理

模型	Long Context Reasoning 长上下文推理
Claude Opus 5.5 Anthropic · Max	84.7%
Claude Sonnet 5.5 Anthropic · Max	82.7%
Muse Wonder Full Meta · Max	83.7%
Claude Fable 5.1 Anthropic · Max	85.3%
GPT-6 Astra OpenAI · Max	80.7%
Gemini 4 Argon Google · High	79.7%
Muse Wonder Meta · Max	80.7%
GPT-6.1 Sol OpenAI · Max	83.0%
Muse Spark 1.3 Meta · Max	83.0%
Grok 4.7 xAI · Xhigh	76.7%
MiMo-V2.6-Pro 小米	86.3%
Qwen3.8 Max 阿里通义 · 0902	80.3%
GLM-5.3 智谱 · Max	79.7%
Step 5 Preview 阶跃星辰	88.3%
Kimi K3 月之暗面 · Max	88.7%
Claude Haiku 5.5 Anthropic · Max	82.7%
GPT-5.6 Terra OpenAI · Max	83.0%
GLM-5.3-Flash 智谱	80.0%
Ling 3.1 Flash InclusionAI	83.0%
Gemini 3.8 Flash Google · High	81.3%
Qwen3.8 2.4T A95B 阿里通义	80.3%
Qwen3.8-Flash-Next 阿里通义	79.7%
DeepSeek V4.1 Flash DeepSeek · Max	84.0%
Muse Spark Open Meta · Xhigh	78.3%
Mistral Large 4 Preview Mistral	81.3%
GPT-6 Luna OpenAI · Max	83.3%
MiMo-V2.6-Flash 小米	74.3%
DeepSeek V4 Pro 0813 DeepSeek · Max	80.3%
Agnes 2.5 Pro Beta Sapiens AI	83.0%
DeepSeek V4 Flash Vision DeepSeek · Max	81.3%
JT-4.1 Flash 236B A21B China Mobile · Reasoning	78.3%
Qwen3.8 27B 阿里通义 · Xhigh	82.0%
Motif 3 Motif Technologies	76.0%
GPT-5.3 Codex OpenAI · Xhigh	83.3%
K2 Horizon 375B A23B Institute of Foundation Models	80.0%
Beam 1.0 Reflection	78.0%

当前难度天花板

模型	CritPt 科研批判性推理
Claude Opus 5.5 Anthropic · Max	31.7%
Claude Sonnet 5.5 Anthropic · Max	31.4%
Muse Wonder Full Meta · Max	32.9%
Claude Fable 5.1 Anthropic · Max	29.7%
GPT-6 Astra OpenAI · Max	31.7%
Gemini 4 Argon Google · High	27.1%
Muse Wonder Meta · Max	28.3%
GPT-6.1 Sol OpenAI · Max	31.7%
Muse Spark 1.3 Meta · Max	24.9%
Grok 4.7 xAI · Xhigh	17.7%
MiMo-V2.6-Pro 小米	26.6%
Qwen3.8 Max 阿里通义 · 0902	17.7%
GLM-5.3 智谱 · Max	19.1%
Step 5 Preview 阶跃星辰	20.9%
Kimi K3 月之暗面 · Max	23.4%
Claude Haiku 5.5 Anthropic · Max	18.9%
GPT-5.6 Terra OpenAI · Max	30.0%
GLM-5.3-Flash 智谱	15.4%
Ling 3.1 Flash InclusionAI	18.0%
Gemini 3.8 Flash Google · High	18.3%
Qwen3.8 2.4T A95B 阿里通义	20.0%
Qwen3.8-Flash-Next 阿里通义	11.1%
DeepSeek V4.1 Flash DeepSeek · Max	14.3%
Muse Spark Open Meta · Xhigh	14.9%
Mistral Large 4 Preview Mistral	10.6%
GPT-6 Luna OpenAI · Max	19.4%
MiMo-V2.6-Flash 小米	12.0%
DeepSeek V4 Pro 0813 DeepSeek · Max	18.0%
Agnes 2.5 Pro Beta Sapiens AI	15.7%
DeepSeek V4 Flash Vision DeepSeek · Max	10.9%
JT-4.1 Flash 236B A21B China Mobile · Reasoning	14.0%
Qwen3.8 27B 阿里通义 · Xhigh	5.4%
Motif 3 Motif Technologies	6.6%
GPT-5.3 Codex OpenAI · Xhigh	16.9%
K2 Horizon 375B A23B Institute of Foundation Models	4.6%
Beam 1.0 Reflection	13.1%

各维度当前 SOTA

入选的 36 个模型里，该维度的最高实测分。「覆盖」= 有实测分的模型数。

AA Intelligence Index

综合智能指数 · AA 统一打分体系

待建卡

57.6 Claude Opus 5.5

Anthropic

覆盖 36/36

Artificial Analysis 用统一协议把多个 benchmark 合成的综合指数，0–100 分，可横向对比

AA Omniscience

全知指数 · 准确率 − 幻觉率

待建卡

+46.4 Claude Opus 5.5

Anthropic

覆盖 36/36

同时看「答对多少」和「编造多少」，负值代表幻觉多于答对

GPQA Diamond

研究生级科学问答 · pass@1

评测卡

96.1% GPT-6 Astra

OpenAI

覆盖 19/36

博士级物理/化学/生物选择题，人类专家正确率约 65%

Humanity's Last Exam

终极考试 · pass@1

评测卡

61.4% Claude Opus 5.5

Anthropic

覆盖 36/36

全球专家贡献的高难度综合题，纯文本无工具，当前最难的通用知识榜之一

MMMU-Pro

多模态理解 Pro · accuracy

评测卡

87.7% Claude Opus 5.5

Anthropic

覆盖 18/36

MMMU 的抗污染加固版，30 学科大学级图文题，纯文本提示无法作答

SciCode

科研编程 · pass@1

评测卡

66.9% Claude Opus 5.5

Anthropic

覆盖 35/36

真实科研场景的代码实现题，覆盖物理/化学/生物/数学

IFBench

指令跟随 · accuracy

待建卡

75.4% GPT-5.3 Codex

OpenAI

覆盖 2/36

可验证指令约束的跟随能力，测「按要求做」而不是「答得对」

τ²-bench

工具调用智能体 · accuracy

评测卡

86.3% GPT-5.6 Terra

OpenAI

覆盖 2/36

双控制（用户 + 工具）的多轮对话智能体评测，主域为零售/航空

τ²-bench (Banking)

金融工具智能体 · accuracy

评测卡

50.5% Muse Spark 1.3

口径说明

同一模型的多个 reasoning 档位只保留综合指数最高的一档，分数即为该档位实测值。
所有分数由 Artificial Analysis 用同一协议复测，不是厂商自报；因此同一 benchmark 的数值可能与厂商 technical report 里的数字不同（例如 Terminal-Bench 4.0），横向对比请以本页口径为准。
标「待建卡」的维度说明本站评测卡目录尚未收录，会在后续补卡后自动挂上链接。