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 · 9月22日Grok 4.7 (Xhigh)评测结果上线模型评测 · 9月22日MiMo-V2.6-Pro评测结果上线社区洞察 · 9月21日“上桌？Step 5 Preview实战能力评测”发布模型评测 · 9月20日Step 5 Preview (High)评测结果上线社区洞察 · 9月20日“GLM-5.3-FlashX，新晋性价比之王”发布社区洞察 · 9月19日“让豆包新模型干了24小时，效果却不如旧模型？”发布模型评测 · 9月19日GLM-5.3-FlashX (Max)评测结果上线模型评测 · 9月19日GLM-5.3-Flash (Max)评测结果上线模型评测 · 9月19日GLM-5.3 (Max)评测结果上线社区洞察 · 9月18日“追踪2142项AI技术动态后，我们发现最值得关注的不是新模型”发布模型评测 · 9月18日Doubao-Seed-2.1-pro-260915 (High)评测结果上线社区洞察 · 9月17日“DeepSeek真能“斩杀”千问和豆包吗？”发布模型评测 · 9月16日Qwen3.8-Max (Xhigh)评测结果上线模型评测 · 9月16日Qwen3.8-Flash (Xhigh)评测结果上线社区洞察 · 9月16日“DeepSeek V4.1 Flash真的超越V4 Pro了吗？”发布社区洞察 · 9月15日“本周AI技术跟踪：失效信息将影响Agent记忆可靠性（W37）”发布模型评测 · 9月15日DeepSeek-V4-Pro-0813 (Max)评测结果上线社区洞察 · 9月14日“研究精选：13份报告，聚焦企业AI的成本、落地与组织能力”发布模型评测 · 9月14日DeepSeek-V4.1-Flash (Max)评测结果上线社区洞察 · 9月11日“英伟达给AI加上长期记忆，Agent要拼的不只是智商了”发布更多日志
SQBENCH RESEARCH · NEW

SQBench 评测基准论文

面向真实业务场景的大模型实战评测框架

阅读论文下载 PDF

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。

排名模型按实战能力原始分排序，同分时按最终表现原始分排序；两项均同分则并列。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
1GPT-6 Astra (Max)OpenAI64.390.0%3.3%$168.962Claude Fable 5.1 (Max)Anthropic63.187.1%3.3%$376.303Kimi K3月之暗面60.588.1%6.2%$42.174Claude Opus 4.8 (Max)Anthropic57.683.3%4.7%$201.225GLM-5.3 (Max)智谱56.179.9%7.2%$29.586MiMo-V2.6-Pro小米54.882.9%5.9%$4.457GPT-5.6 Sol (Max)OpenAI54.685.7%4.3%$103.148GLM-5.3-FlashX (Max)智谱54.680.4%6.8%$5.689GLM-5.2 (Max)智谱52.788.3%6.2%$27.0410Gemini 3.6 Flash (High)Google51.882.6%4.2%$81.77

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。

64.3GPT-6 Astra (Max)
63.1Claude Fable 5.1 (Max)
60.5Kimi K3
57.6Claude Opus 4.8 (Max)
56.1GLM-5.3 (Max)
54.8MiMo-V2.6-Pro
54.6GPT-5.6 Sol (Max)
54.6GLM-5.3-FlashX (Max)
52.7GLM-5.2 (Max)
51.8Gemini 3.6 Flash (High)
51.4DeepSeek-V4-Pro-0813 (Max)
51.3Doubao-Seed-2.1-turbo (High)
51.3Qwen3.8-Max (Xhigh)
51.1Qwen3.7-Max
50.6Grok 4.7 (Xhigh)
50.4Kimi K2.7 Code
50.0Step 5 Preview (High)
49.7GLM-5.3-Flash (Max)
49.5DeepSeek-V4-Pro (Max)
48.5MiMo-V2.5
48.3DeepSeek-V4-Pro (High)
48.1Doubao-Seed-2.1-pro (High)
48.0DeepSeek-V4.1-Flash (Max)
47.3Kimi K2.6
47.1DeepSeek-V4-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
64.3GPT-6 Astra (Max)
63.1Claude Fable 5.1 (Max)
60.5Kimi K3
57.6Claude Opus 4.8 (Max)
56.1GLM-5.3 (Max)
54.8MiMo-V2.6-Pro
54.6GPT-5.6 Sol (Max)
54.6GLM-5.3-FlashX (Max)
52.7GLM-5.2 (Max)
51.8Gemini 3.6 Flash (High)
51.4DeepSeek-V4-Pro-0813 (Max)
51.3Doubao-Seed-2.1-turbo (High)
51.3Qwen3.8-Max (Xhigh)
51.1Qwen3.7-Max
50.6Grok 4.7 (Xhigh)
50.4Kimi K2.7 Code
50.0Step 5 Preview (High)
49.7GLM-5.3-Flash (Max)
49.5DeepSeek-V4-Pro (Max)
48.5MiMo-V2.5
48.3DeepSeek-V4-Pro (High)
48.1Doubao-Seed-2.1-pro (High)
48.0DeepSeek-V4.1-Flash (Max)
47.3Kimi K2.6
47.1DeepSeek-V4-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

68.0GPT-6 Astra (Max)
64.0Claude Fable 5.1 (Max)
56.0Kimi K3
56.0GLM-5.3 (Max)
56.0GLM-5.2 (Max)
56.0Qwen3.7-Max
56.0DeepSeek-V4.1-Flash (Max)
52.0Claude Opus 4.8 (Max)
52.0GPT-5.6 Sol (Max)
52.0DeepSeek-V4-Pro-0813 (Max)
52.0Qwen3.8-Max (Xhigh)
52.0Grok 4.7 (Xhigh)
48.0GLM-5.3-FlashX (Max)
48.0Gemini 3.6 Flash (High)
48.0Doubao-Seed-2.1-turbo (High)
48.0MiMo-V2.5
40.0MiMo-V2.6-Pro
40.0Kimi K2.7 Code
40.0GLM-5.3-Flash (Max)
36.0DeepSeek-V4-Pro (Max)
36.0Kimi K2.6
32.0Step 5 Preview (High)
32.0DeepSeek-V4-Pro (High)
28.0Doubao-Seed-2.1-pro (High)
20.0DeepSeek-V4-Flash (High)

长文本推理

越高越好

92.0GLM-5.3 (Max)
64.0Claude Fable 5.1 (Max)
64.0Kimi K3
64.0GLM-5.3-Flash (Max)
60.0DeepSeek-V4.1-Flash (Max)
56.0MiMo-V2.6-Pro
52.0GLM-5.3-FlashX (Max)
52.0Step 5 Preview (High)
48.0Claude Opus 4.8 (Max)
48.0DeepSeek-V4-Pro-0813 (Max)
48.0Qwen3.8-Max (Xhigh)
44.0Grok 4.7 (Xhigh)
40.0GPT-6 Astra (Max)
40.0GLM-5.2 (Max)
32.0Kimi K2.7 Code
32.0MiMo-V2.5
28.0Qwen3.7-Max
28.0Doubao-Seed-2.1-pro (High)
24.0Gemini 3.6 Flash (High)
24.0Kimi K2.6
20.0GPT-5.6 Sol (Max)
20.0DeepSeek-V4-Pro (Max)
20.0DeepSeek-V4-Flash (High)
16.0Doubao-Seed-2.1-turbo (High)
8.0DeepSeek-V4-Pro (High)

动态代码逻辑

越高越好

100.0GPT-6 Astra (Max)
100.0Claude Fable 5.1 (Max)
100.0Qwen3.8-Max (Xhigh)
100.0Grok 4.7 (Xhigh)
96.0Claude Opus 4.8 (Max)
96.0GPT-5.6 Sol (Max)
96.0GLM-5.3-FlashX (Max)
96.0Gemini 3.6 Flash (High)
96.0DeepSeek-V4.1-Flash (Max)
92.0Doubao-Seed-2.1-turbo (High)
92.0Step 5 Preview (High)
88.0Kimi K3
88.0MiMo-V2.6-Pro
88.0Qwen3.7-Max
88.0DeepSeek-V4-Pro (Max)
84.0GLM-5.3 (Max)
76.0DeepSeek-V4-Pro-0813 (Max)
76.0DeepSeek-V4-Pro (High)
72.0Kimi K2.6
68.0GLM-5.2 (Max)
68.0GLM-5.3-Flash (Max)
64.0Doubao-Seed-2.1-pro (High)
60.0Kimi K2.7 Code
56.0DeepSeek-V4-Flash (High)
52.0MiMo-V2.5

鲁棒性与边界

越高越好

68.0Claude Fable 5.1 (Max)
64.0GPT-6 Astra (Max)
64.0GPT-5.6 Sol (Max)
56.0Claude Opus 4.8 (Max)
52.0DeepSeek-V4-Pro-0813 (Max)
48.0Gemini 3.6 Flash (High)
48.0DeepSeek-V4.1-Flash (Max)
44.0Step 5 Preview (High)
44.0DeepSeek-V4-Pro (High)
40.0Qwen3.8-Max (Xhigh)
40.0DeepSeek-V4-Pro (Max)
36.0Kimi K3
36.0GLM-5.3 (Max)
36.0GLM-5.3-FlashX (Max)
36.0GLM-5.2 (Max)
36.0Qwen3.7-Max
36.0Grok 4.7 (Xhigh)
36.0Kimi K2.7 Code
36.0GLM-5.3-Flash (Max)
36.0Doubao-Seed-2.1-pro (High)
32.0MiMo-V2.6-Pro
32.0MiMo-V2.5
32.0DeepSeek-V4-Flash (High)
28.0Kimi K2.6
24.0Doubao-Seed-2.1-turbo (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。

90.0%GPT-6 Astra (Max)
88.3%GLM-5.2 (Max)
88.1%Kimi K3
87.5%Kimi K2.6
87.1%Claude Fable 5.1 (Max)
86.3%Qwen3.7-Max
85.8%Doubao-Seed-2.1-turbo (High)
85.7%GPT-5.6 Sol (Max)
84.9%Doubao-Seed-2.1-pro (High)
84.2%Kimi K2.7 Code
83.7%DeepSeek-V4-Pro-0813 (Max)
83.3%Claude Opus 4.8 (Max)
83.2%DeepSeek-V4-Flash (High)
82.9%MiMo-V2.6-Pro
82.6%Gemini 3.6 Flash (High)
82.2%DeepSeek-V4-Pro (High)
81.0%Grok 4.7 (Xhigh)
80.4%GLM-5.3-FlashX (Max)
79.9%GLM-5.3 (Max)
79.9%DeepSeek-V4-Pro (Max)
79.0%MiMo-V2.5
78.3%GLM-5.3-Flash (Max)
77.1%DeepSeek-V4.1-Flash (Max)
75.8%Qwen3.8-Max (Xhigh)
75.7%Step 5 Preview (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)
$1.40MiMo-V2.5$1.40输入$0.70推理$0.39回答$0.31MiMo-V2.5
$2.27GLM-5.3-Flash (Max)$2.27输入$1.13推理$0.67回答$0.47GLM-5.3-Flash (Max)
$2.87DeepSeek-V4-Pro (High)$2.87输入$1.18推理$0.78回答$0.91DeepSeek-V4-Pro (High)
$3.98DeepSeek-V4-Pro (Max)$3.98输入$1.93推理$1.08回答$0.97DeepSeek-V4-Pro (Max)
$4.45MiMo-V2.6-Pro$4.45输入$1.90推理$1.55回答$1.00MiMo-V2.6-Pro
$5.21DeepSeek-V4.1-Flash (Max)$5.21输入$0.68推理$3.73回答$0.80DeepSeek-V4.1-Flash (Max)
$5.68GLM-5.3-FlashX (Max)$5.68输入$2.79推理$1.69回答$1.20GLM-5.3-FlashX (Max)
$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)
$9.83DeepSeek-V4-Pro-0813 (Max)$9.83输入$2.10推理$5.53回答$2.21DeepSeek-V4-Pro-0813 (Max)
$14.98Kimi K2.7 Code$14.98输入$7.08推理$4.92回答$2.99Kimi K2.7 Code
$16.72Step 5 Preview (High)$16.72输入$8.78推理$0.00回答$7.94Step 5 Preview (High)
$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)
$27.04GLM-5.2 (Max)$27.04输入$12.97推理$9.92回答$4.15GLM-5.2 (Max)
$29.58GLM-5.3 (Max)$29.58输入$14.06推理$10.45回答$5.07GLM-5.3 (Max)
$30.13Qwen3.7-Max$30.13输入$19.67推理$5.44回答$5.02Qwen3.7-Max
$38.18Qwen3.8-Max (Xhigh)$38.18输入$19.38推理$9.19回答$9.61Qwen3.8-Max (Xhigh)
$42.17Kimi K3$42.17输入$13.19推理$16.45回答$12.54Kimi K3
$50.80$31.65Grok 4.7 (Xhigh)$50.80输入$31.65推理$13.35回答$5.80Grok 4.7 (Xhigh)
$81.77$51.93$20.96Gemini 3.6 Flash (High)$81.77输入$51.93推理$20.96回答$8.88Gemini 3.6 Flash (High)
$103.14$41.71$37.89$23.53GPT-5.6 Sol (Max)$103.14输入$41.71推理$37.89回答$23.53GPT-5.6 Sol (Max)
$168.96$73.65$60.01$35.30GPT-6 Astra (Max)$168.96输入$73.65推理$60.01回答$35.30GPT-6 Astra (Max)
$201.22$70.92$116.02Claude Opus 4.8 (Max)$201.22输入$70.92推理$14.28回答$116.02Claude Opus 4.8 (Max)
$376.30$89.59$172.76$113.95Claude Fable 5.1 (Max)$376.30输入$89.59推理$172.76回答$113.95Claude Fable 5.1 (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。

OpenAIAnthropic月之暗面智谱小米GoogleDeepSeek字节跳动阿里巴巴xAI阶跃星辰
理想区域实战能力可靠度50.055.060.065.075.0%80.0%85.0%90.0%GPT-6 Astra (Max)GPT-6 Astra (Max)可靠度90.0%实战能力64.3%Click for more infoClaude Fable 5.1 (Max)Claude Fable 5.1 (Max)可靠度87.1%实战能力63.1%Click for more infoKimi K3Kimi K3可靠度88.1%实战能力60.5%Click for more infoClaude Opus 4.8 (Max)Claude Opus 4.8 (Max)可靠度83.3%实战能力57.6%Click for more infoGLM-5.3 (Max)GLM-5.3 (Max)可靠度79.9%实战能力56.1%Click for more infoMiMo-V2.6-ProMiMo-V2.6-Pro可靠度82.9%实战能力54.8%Click for more infoGPT-5.6 Sol (Max)GPT-5.6 Sol (Max)可靠度85.7%实战能力54.6%Click for more infoGLM-5.3-FlashX (Max)GLM-5.3-FlashX (Max)可靠度80.4%实战能力54.6%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 infoGemini 3.6 Flash (High)Gemini 3.6 Flash (High)可靠度82.6%实战能力51.8%Click for more infoDeepSeek-V4-Pro-0813 (Max)DeepSeek-V4-Pro-0813 (Max)可靠度83.7%实战能力51.4%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 infoQwen3.8-Max (Xhigh)Qwen3.8-Max (Xhigh)可靠度75.8%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 infoGrok 4.7 (Xhigh)Grok 4.7 (Xhigh)可靠度81.0%实战能力50.6%Click for more infoKimi K2.7 CodeKimi K2.7 Code可靠度84.2%实战能力50.4%Click for more infoStep 5 Preview (High)Step 5 Preview (High)可靠度75.7%实战能力50.0%Click for more infoGLM-5.3-Flash (Max)GLM-5.3-Flash (Max)可靠度78.3%实战能力49.7%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 infoDeepSeek-V4.1-Flash (Max)DeepSeek-V4.1-Flash (Max)可靠度77.1%实战能力48.0%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 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)
0.030.120.41
GLM-5.3-Flash (Max)缓存$0.03输入$0.12输出$0.41GLM-5.3-Flash (Max)
0.00300.150.60
DeepSeek-V4.1-Flash (Max)缓存$0.0030输入$0.15输出$0.60DeepSeek-V4.1-Flash (Max)
0.00360.440.87
MiMo-V2.6-Pro缓存$0.0036输入$0.44输出$0.87MiMo-V2.6-Pro
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)
0.080.291.03
GLM-5.3-FlashX (Max)缓存$0.08输入$0.29输出$1.03GLM-5.3-FlashX (Max)
0.020.661.98
DeepSeek-V4-Pro-0813 (Max)缓存$0.02输入$0.66输出$1.98DeepSeek-V4-Pro-0813 (Max)
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)
0.051.032.94
Step 5 Preview (High)缓存$0.05输入$1.03输出$2.94Step 5 Preview (High)
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-5.3 (Max)缓存$0.29输入$1.18输出$4.12GLM-5.3 (Max)
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)
0.401.604.80
Grok 4.7 (Xhigh)缓存$0.40输入$1.60输出$4.80Grok 4.7 (Xhigh)
0.221.775.29
Qwen3.8-Max (Xhigh)缓存$0.22输入$1.77输出$5.29Qwen3.8-Max (Xhigh)
0.351.775.29
Qwen3.7-Max缓存$0.35输入$1.77输出$5.29Qwen3.7-Max
0.151.507.50
Gemini 3.6 Flash (High)缓存$0.15输入$1.50输出$7.50Gemini 3.6 Flash (High)
0.292.9414.71
Kimi K3缓存$0.29输入$2.94输出$14.71Kimi K3
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)
0.505.0030.00
GPT-5.6 Sol (Max)缓存$0.50输入$5.00输出$30.00GPT-5.6 Sol (Max)
1.0010.0050.00
GPT-6 Astra (Max)缓存$1.00输入$10.00输出$50.00GPT-6 Astra (Max)
0.2510.0050.00
Claude Fable 5.1 (Max)缓存$0.25输入$10.00输出$50.00Claude Fable 5.1 (Max)

推理模型以此图标标识