沙丘社区

SQBench

大模型实战能力评测基准

沙丘社区是中国领先的AI研究与评测机构。SQBench聚焦大模型在真实业务场景中的实战表现，围绕任务交付、风险边界与成本效率等维度，为模型选型、能力对比和部署决策提供参考依据。

查看实战榜单阅读评测方法

更新日志

模型评测 · 7月8日DeepSeek-V4-Pro (Max)评测结果上线模型评测 · 7月8日MiniMax-M2.7评测结果上线模型评测 · 7月8日Qwen3.6-Plus评测结果上线模型评测 · 7月8日Qwen3.7-Max评测结果上线模型评测 · 7月8日Qwen3.7-Plus评测结果上线模型评测 · 7月8日Gemini 3.5 Flash (High)评测结果上线模型评测 · 7月8日Hy3评测结果上线模型评测 · 7月8日MiMo-V2.5评测结果上线模型评测 · 7月8日MiMo-V2.5-Pro评测结果上线模型评测 · 7月8日GLM-5.1评测结果上线模型评测 · 7月8日GLM-5.2 (Max)评测结果上线模型评测 · 7月8日Doubao-Seed-2.1-turbo (High)评测结果上线模型评测 · 7月8日DeepSeek-V4-Flash (Max)评测结果上线模型评测 · 7月8日DeepSeek-V4-Pro (High)评测结果上线模型评测 · 7月8日Kimi K2.6评测结果上线模型评测 · 7月8日Kimi K2.7 Code评测结果上线模型评测 · 7月8日Step 3.5 Flash (High)评测结果上线模型评测 · 7月8日Step 3.7 Flash (High)评测结果上线模型评测 · 7月8日Doubao-Seed-2.1-pro (High)评测结果上线模型评测 · 7月8日DeepSeek-V4-Flash (High)评测结果上线更多日志

Leaderboard

实战能力排行榜

基于 SQBench 实战任务结果排序，展示模型在真实业务场景中的综合表现。

排名模型满分 100，越高表示实战能力越强未触发风险扣分的任务比例，越高越好触发幻觉相关风险的任务比例，越低越好完成本次评测的模型 API 成本
1Claude Opus 4.8 (Max)Anthropic57.683.3%4.7%$201.222GLM-5.2 (Max)智谱52.788.3%6.2%$27.043Doubao-Seed-2.1-turbo (High)字节跳动51.385.8%7.3%$8.744Qwen3.7-Max阿里巴巴51.186.3%5.5%$30.135Kimi K2.7 Code月之暗面50.484.2%4.9%$14.986DeepSeek-V4-Pro (Max)DeepSeek49.579.9%4.2%$3.987MiMo-V2.5小米48.579.0%6.4%$1.408DeepSeek-V4-Pro (High)DeepSeek48.382.2%6.2%$2.879Doubao-Seed-2.1-pro (High)字节跳动48.184.9%7.7%$21.5910Kimi K2.6月之暗面47.387.5%4.2%$20.42

Models

模型表现

展示模型在 SQBench 实战任务中的总体表现、完成质量与风险影响。

实战能力

基于任务通过情况衡量模型在真实业务任务中的稳定交付表现，越高表示可稳定完成的任务比例越高。

57.6Claude Opus 4.8 (Max)
52.7GLM-5.2 (Max)
51.3Doubao-Seed-2.1-turbo (High)
51.1Qwen3.7-Max
50.4Kimi K2.7 Code
49.5DeepSeek-V4-Pro (Max)
48.5MiMo-V2.5
48.3DeepSeek-V4-Pro (High)
48.1Doubao-Seed-2.1-pro (High)
47.3Kimi K2.6
47.1DeepSeek-V4-Flash (High)
46.7GLM-5.1
46.7MiniMax-M3
45.4Qwen3.6-Plus
45.3Qwen3.7-Plus
45.3Hy3
44.0MiMo-V2.5-Pro
40.6Hy3 Preview
40.5Gemini 3.5 Flash (High)
37.9MiniMax-M2.7
37.4Grok 4.3 (High)
36.8DeepSeek-V4-Flash (Max)
34.1Gemma 4 31B
28.4Step 3.5 Flash (High)
18.7Step 3.7 Flash (High)

推理模型以此图标标识

实战能力

实战能力采用任务通过口径计算，强调模型在标准评测条件下稳定达标的能力；最终表现分和任务完成分保留连续得分，用于辅助分析任务完成质量与风险影响。

实战能力（按开放属性划分）

按开源/闭源展示模型实战能力，用于观察不同开放属性模型的表现分布。

闭源模型开源模型
57.6Claude Opus 4.8 (Max)
52.7GLM-5.2 (Max)
51.3Doubao-Seed-2.1-turbo (High)
51.1Qwen3.7-Max
50.4Kimi K2.7 Code
49.5DeepSeek-V4-Pro (Max)
48.5MiMo-V2.5
48.3DeepSeek-V4-Pro (High)
48.1Doubao-Seed-2.1-pro (High)
47.3Kimi K2.6
47.1DeepSeek-V4-Flash (High)
46.7GLM-5.1
46.7MiniMax-M3
45.4Qwen3.6-Plus
45.3Qwen3.7-Plus
45.3Hy3
44.0MiMo-V2.5-Pro
40.6Hy3 Preview
40.5Gemini 3.5 Flash (High)
37.9MiniMax-M2.7
37.4Grok 4.3 (High)
36.8DeepSeek-V4-Flash (Max)
34.1Gemma 4 31B
28.4Step 3.5 Flash (High)
18.7Step 3.7 Flash (High)

推理模型以此图标标识

Capability

实战能力分析

从基础执行、复合工作流和业务场景三个层面，比较模型在不同类型实战任务中的表现。

L1 原子能力

评估模型在基础认知、指令执行、长上下文推理、代码逻辑和边界处理中的稳定性。

指令遵循

越高越好

56.0GLM-5.2 (Max)
56.0Qwen3.7-Max
52.0Claude Opus 4.8 (Max)
52.0GLM-5.1
48.0Doubao-Seed-2.1-turbo (High)
48.0MiMo-V2.5
48.0Qwen3.7-Plus
44.0Qwen3.6-Plus
44.0Hy3 Preview
44.0Gemini 3.5 Flash (High)
40.0Kimi K2.7 Code
36.0DeepSeek-V4-Pro (Max)
36.0Kimi K2.6
36.0MiniMax-M3
36.0Gemma 4 31B
32.0DeepSeek-V4-Pro (High)
32.0Hy3
32.0Grok 4.3 (High)
32.0Step 3.7 Flash (High)
28.0Doubao-Seed-2.1-pro (High)
28.0MiMo-V2.5-Pro
20.0DeepSeek-V4-Flash (High)
20.0DeepSeek-V4-Flash (Max)
20.0Step 3.5 Flash (High)
12.0MiniMax-M2.7

长文本推理

越高越好

48.0Claude Opus 4.8 (Max)
40.0GLM-5.2 (Max)
36.0Hy3
32.0Kimi K2.7 Code
32.0MiMo-V2.5
28.0Qwen3.7-Max
28.0Doubao-Seed-2.1-pro (High)
24.0Kimi K2.6
24.0GLM-5.1
24.0MiniMax-M3
24.0Qwen3.7-Plus
20.0DeepSeek-V4-Pro (Max)
20.0DeepSeek-V4-Flash (High)
20.0MiMo-V2.5-Pro
16.0Doubao-Seed-2.1-turbo (High)
16.0Hy3 Preview
16.0Gemini 3.5 Flash (High)
16.0MiniMax-M2.7
12.0Qwen3.6-Plus
12.0Step 3.5 Flash (High)
8.0DeepSeek-V4-Pro (High)
8.0Grok 4.3 (High)
8.0DeepSeek-V4-Flash (Max)
4.0Gemma 4 31B
4.0Step 3.7 Flash (High)

动态代码逻辑

越高越好

96.0Claude Opus 4.8 (Max)
92.0Doubao-Seed-2.1-turbo (High)
92.0Gemma 4 31B
88.0Qwen3.7-Max
88.0DeepSeek-V4-Pro (Max)
88.0Gemini 3.5 Flash (High)
88.0Grok 4.3 (High)
84.0Qwen3.7-Plus
80.0Qwen3.6-Plus
76.0DeepSeek-V4-Pro (High)
72.0Kimi K2.6
72.0Hy3
72.0DeepSeek-V4-Flash (Max)
68.0GLM-5.2 (Max)
68.0Hy3 Preview
64.0Doubao-Seed-2.1-pro (High)
60.0Kimi K2.7 Code
60.0GLM-5.1
56.0DeepSeek-V4-Flash (High)
56.0MiniMax-M3
56.0Step 3.7 Flash (High)
52.0MiMo-V2.5
48.0MiMo-V2.5-Pro
48.0MiniMax-M2.7
40.0Step 3.5 Flash (High)

鲁棒性与边界

越高越好

56.0Claude Opus 4.8 (Max)
48.0Gemini 3.5 Flash (High)
44.0DeepSeek-V4-Pro (High)
40.0DeepSeek-V4-Pro (Max)
36.0GLM-5.2 (Max)
36.0Qwen3.7-Max
36.0Kimi K2.7 Code
36.0Doubao-Seed-2.1-pro (High)
36.0DeepSeek-V4-Flash (Max)
32.0MiMo-V2.5
32.0DeepSeek-V4-Flash (High)
32.0GLM-5.1
32.0Qwen3.6-Plus
32.0Hy3
28.0Kimi K2.6
28.0MiniMax-M2.7
24.0Doubao-Seed-2.1-turbo (High)
24.0MiniMax-M3
24.0Qwen3.7-Plus
24.0MiMo-V2.5-Pro
24.0Hy3 Preview
20.0Grok 4.3 (High)
16.0Gemma 4 31B
16.0Step 3.5 Flash (High)
16.0Step 3.7 Flash (High)

推理模型以此图标标识

Safety & Reliability

安全与合规

对比模型在任务执行中的可靠性、风险控制和幻觉相关表现。

可靠度

未触发风险扣分的任务比例，越高表示执行过程越稳定。

88.3%GLM-5.2 (Max)
88.3%Grok 4.3 (High)
87.5%Kimi K2.6
86.7%GLM-5.1
86.3%Qwen3.7-Max
85.8%Doubao-Seed-2.1-turbo (High)
84.9%Doubao-Seed-2.1-pro (High)
84.9%Gemma 4 31B
84.2%Kimi K2.7 Code
83.6%Hy3
83.3%Claude Opus 4.8 (Max)
83.2%DeepSeek-V4-Flash (High)
83.0%Qwen3.6-Plus
82.7%MiniMax-M3
82.7%Qwen3.7-Plus
82.2%DeepSeek-V4-Pro (High)
80.5%Hy3 Preview
80.2%MiniMax-M2.7
79.9%DeepSeek-V4-Pro (Max)
79.3%MiMo-V2.5-Pro
79.0%MiMo-V2.5
74.6%Gemini 3.5 Flash (High)
73.6%Step 3.5 Flash (High)
73.3%DeepSeek-V4-Flash (Max)
69.1%Step 3.7 Flash (High)

推理模型以此图标标识

Cost Efficiency

成本效率

展示模型完成 SQBench 实战任务所需的评测成本、完成耗时和 Token 消耗，用于辅助判断规模化使用成本。

评测成本

完成本次评测产生的模型 API 成本，越低表示调用成本越低。

输入推理回答
$1.15DeepSeek-V4-Flash (High)$1.15输入$0.54推理$0.29回答$0.31DeepSeek-V4-Flash (High)
$1.23DeepSeek-V4-Flash (Max)$1.23输入$0.49推理$0.41回答$0.33DeepSeek-V4-Flash (Max)
$1.40MiMo-V2.5$1.40输入$0.70推理$0.39回答$0.31MiMo-V2.5
$1.44Hy3$1.44输入$0.76推理$0.00回答$0.68Hy3
$1.50Gemma 4 31B$1.50输入$1.25推理$0.00回答$0.25Gemma 4 31B
$1.86Step 3.5 Flash (High)$1.86输入$1.14推理$0.00回答$0.72Step 3.5 Flash (High)
$2.87DeepSeek-V4-Pro (High)$2.87输入$1.18推理$0.78回答$0.91DeepSeek-V4-Pro (High)
$3.81MiMo-V2.5-Pro$3.81输入$1.98推理$0.89回答$0.95MiMo-V2.5-Pro
$3.98DeepSeek-V4-Pro (Max)$3.98输入$1.93推理$1.08回答$0.97DeepSeek-V4-Pro (Max)
$4.38Step 3.7 Flash (High)$4.38输入$1.52推理$0.00回答$2.85Step 3.7 Flash (High)
$5.41MiniMax-M2.7$5.41输入$2.83推理$0.54回答$2.04MiniMax-M2.7
$5.49Hy3 Preview$5.49输入$4.35推理$0.00回答$1.15Hy3 Preview
$5.63MiniMax-M3$5.63输入$2.62推理$1.06回答$1.96MiniMax-M3
$6.10Qwen3.7-Plus$6.10输入$2.58推理$1.83回答$1.70Qwen3.7-Plus
$8.74Doubao-Seed-2.1-turbo (High)$8.74输入$3.08推理$1.92回答$3.75Doubao-Seed-2.1-turbo (High)
$12.18Grok 4.3 (High)$12.18输入$8.38推理$2.72回答$1.08Grok 4.3 (High)
$14.82$11.19Qwen3.6-Plus$14.82输入$11.19推理$1.60回答$2.04Qwen3.6-Plus
$14.87GLM-5.1$14.87输入$8.16推理$3.64回答$3.06GLM-5.1
$14.98Kimi K2.7 Code$14.98输入$7.08推理$4.92回答$2.99Kimi K2.7 Code
$20.42Kimi K2.6$20.42输入$9.42推理$4.34回答$6.66Kimi K2.6
$21.59Doubao-Seed-2.1-pro (High)$21.59输入$8.28推理$4.38回答$8.93Doubao-Seed-2.1-pro (High)
$27.04$12.97GLM-5.2 (Max)$27.04输入$12.97推理$9.92回答$4.15GLM-5.2 (Max)
$30.13$19.67Qwen3.7-Max$30.13输入$19.67推理$5.44回答$5.02Qwen3.7-Max
$74.02$40.78$23.46Gemini 3.5 Flash (High)$74.02输入$40.78推理$23.46回答$9.77Gemini 3.5 Flash (High)
$201.22$70.92$14.28$116.02Claude Opus 4.8 (Max)$201.22输入$70.92推理$14.28回答$116.02Claude Opus 4.8 (Max)

推理模型以此图标标识

Value

性价比

对比模型实战能力与可靠度、评测成本、完成耗时和 Token 消耗，辅助判断不同模型的部署价值。

实战能力 vs 可靠度

对比模型的任务交付表现与风险稳定性，右上区域更优。

Anthropic智谱字节跳动阿里巴巴月之暗面DeepSeek小米MiniMax腾讯GooglexAI阶跃星辰
理想区域实战能力可靠度20.030.040.050.060.070.0%75.0%80.0%85.0%Claude Opus 4.8 (Max)Claude Opus 4.8 (Max)可靠度83.3%实战能力57.6%Click for more infoGLM-5.2 (Max)GLM-5.2 (Max)可靠度88.3%实战能力52.7%Click for more infoDoubao-Seed-2.1-turbo (High)Doubao-Seed-2.1-turbo (High)可靠度85.8%实战能力51.3%Click for more infoQwen3.7-MaxQwen3.7-Max可靠度86.3%实战能力51.1%Click for more infoKimi K2.7 CodeKimi K2.7 Code可靠度84.2%实战能力50.4%Click for more infoDeepSeek-V4-Pro (Max)DeepSeek-V4-Pro (Max)可靠度79.9%实战能力49.5%Click for more infoMiMo-V2.5MiMo-V2.5可靠度79.0%实战能力48.5%Click for more infoDeepSeek-V4-Pro (High)DeepSeek-V4-Pro (High)可靠度82.2%实战能力48.3%Click for more infoDoubao-Seed-2.1-pro (High)Doubao-Seed-2.1-pro (High)可靠度84.9%实战能力48.1%Click for more infoKimi K2.6Kimi K2.6可靠度87.5%实战能力47.3%Click for more infoDeepSeek-V4-Flash (High)DeepSeek-V4-Flash (High)可靠度83.2%实战能力47.1%Click for more infoGLM-5.1GLM-5.1可靠度86.7%实战能力46.7%Click for more infoMiniMax-M3MiniMax-M3可靠度82.7%实战能力46.7%Click for more infoQwen3.6-PlusQwen3.6-Plus可靠度83.0%实战能力45.4%Click for more infoQwen3.7-PlusQwen3.7-Plus可靠度82.7%实战能力45.3%Click for more infoHy3Hy3可靠度83.6%实战能力45.3%Click for more infoMiMo-V2.5-ProMiMo-V2.5-Pro可靠度79.3%实战能力44.0%Click for more infoHy3 PreviewHy3 Preview可靠度80.5%实战能力40.6%Click for more infoGemini 3.5 Flash (High)Gemini 3.5 Flash (High)可靠度74.6%实战能力40.5%Click for more infoMiniMax-M2.7MiniMax-M2.7可靠度80.2%实战能力37.9%Click for more infoGrok 4.3 (High)Grok 4.3 (High)可靠度88.3%实战能力37.4%Click for more infoDeepSeek-V4-Flash (Max)DeepSeek-V4-Flash (Max)可靠度73.3%实战能力36.8%Click for more infoGemma 4 31BGemma 4 31B可靠度84.9%实战能力34.1%Click for more infoStep 3.5 Flash (High)Step 3.5 Flash (High)可靠度73.6%实战能力28.4%Click for more infoStep 3.7 Flash (High)Step 3.7 Flash (High)可靠度69.1%实战能力18.7%Click for more info

推理模型以此图标标识

Price

价格

展示模型公开调用单价，用于理解不同模型的基础计费差异；实际评测成本以成本效率部分为准。

分项价格

展示每百万 Token 的输入、输出和缓存单价，越低越好。

缓存输入输出
0.00280.140.28
MiMo-V2.5缓存$0.0028输入$0.14输出$0.28MiMo-V2.5
0.00280.140.28
DeepSeek-V4-Flash (High)缓存$0.0028输入$0.14输出$0.28DeepSeek-V4-Flash (High)
0.00280.140.28
DeepSeek-V4-Flash (Max)缓存$0.0028输入$0.14输出$0.28DeepSeek-V4-Flash (Max)
0.020.100.31
Step 3.5 Flash (High)缓存$0.02输入$0.10输出$0.31Step 3.5 Flash (High)
0.090.120.35
Gemma 4 31B缓存$0.09输入$0.12输出$0.35Gemma 4 31B
0.040.150.59
Hy3缓存$0.04输入$0.15输出$0.59Hy3
0.090.210.70
Hy3 Preview缓存$0.09输入$0.21输出$0.70Hy3 Preview
0.00360.440.87
DeepSeek-V4-Pro (Max)缓存$0.0036输入$0.44输出$0.87DeepSeek-V4-Pro (Max)
0.00360.440.87
DeepSeek-V4-Pro (High)缓存$0.0036输入$0.44输出$0.87DeepSeek-V4-Pro (High)
0.00360.440.87
MiMo-V2.5-Pro缓存$0.0036输入$0.44输出$0.87MiMo-V2.5-Pro
0.040.201.19
Step 3.7 Flash (High)缓存$0.04输入$0.20输出$1.19Step 3.7 Flash (High)
0.060.311.24
MiniMax-M3缓存$0.06输入$0.31输出$1.24MiniMax-M3
0.060.311.24
MiniMax-M2.7缓存$0.06输入$0.31输出$1.24MiniMax-M2.7
0.030.291.77
Qwen3.6-Plus缓存$0.03输入$0.29输出$1.77Qwen3.6-Plus
0.030.291.77
Qwen3.7-Plus缓存$0.03输入$0.29输出$1.77Qwen3.7-Plus
0.090.442.21
Doubao-Seed-2.1-turbo (High)缓存$0.09输入$0.44输出$2.21Doubao-Seed-2.1-turbo (High)
0.201.252.50
Grok 4.3 (High)缓存$0.20输入$1.25输出$2.50Grok 4.3 (High)
0.190.883.53
GLM-5.1缓存$0.19输入$0.88输出$3.53GLM-5.1
0.190.963.97
Kimi K2.7 Code缓存$0.19输入$0.96输出$3.97Kimi K2.7 Code
0.160.963.97
Kimi K2.6缓存$0.16输入$0.96输出$3.97Kimi K2.6
0.291.184.12
GLM-5.2 (Max)缓存$0.29输入$1.18输出$4.12GLM-5.2 (Max)
0.180.884.41
Doubao-Seed-2.1-pro (High)缓存$0.18输入$0.88输出$4.41Doubao-Seed-2.1-pro (High)
0.351.775.29
Qwen3.7-Max缓存$0.35输入$1.77输出$5.29Qwen3.7-Max
0.151.509.00
Gemini 3.5 Flash (High)缓存$0.15输入$1.50输出$9.00Gemini 3.5 Flash (High)
0.505.0025.00
Claude Opus 4.8 (Max)缓存$0.50输入$5.00输出$25.00Claude Opus 4.8 (Max)

推理模型以此图标标识