沙丘社区

垂直行业 · 政务

Leaderboard

政务行业实战能力排行榜

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

排名模型满分 100，越高表示政务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
1DeepSeek-V4-Pro (Max)DeepSeek42.691.8%1.6%$3.202Claude Opus 4.8 (Max)Anthropic42.386.0%5.5%$139.593MiMo-V2.5小米42.187.0%2.7%$1.114DeepSeek-V4-Pro (High)DeepSeek42.092.6%2.6%$2.245DeepSeek-V4-Flash (High)DeepSeek41.293.6%1.6%$0.906Qwen3.7-Max阿里巴巴39.790.3%6.1%$24.227GLM-5.2 (Max)智谱39.587.5%6.1%$21.138Kimi K2.6月之暗面37.595.4%1.6%$15.839GLM-5.1智谱37.295.2%1.1%$11.5010Hy3腾讯36.889.8%5.2%$1.13

Models

政务行业能力表现

展示模型在 SQBench 政务实战任务中的总体表现、完成质量与风险影响。

政务实战能力

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。

42.6DeepSeek-V4-Pro (Max)
42.3Claude Opus 4.8 (Max)
42.1MiMo-V2.5
42.0DeepSeek-V4-Pro (High)
41.2DeepSeek-V4-Flash (High)
39.7Qwen3.7-Max
39.5GLM-5.2 (Max)
37.5Kimi K2.6
37.2GLM-5.1
36.8Hy3
36.5Qwen3.7-Plus
35.5Doubao-Seed-2.1-turbo (High)
35.2Kimi K2.7 Code
34.4Gemini 3.5 Flash (High)
33.1MiniMax-M2.7
32.0MiMo-V2.5-Pro
30.8Hy3 Preview
30.4Doubao-Seed-2.1-pro (High)
29.7Grok 4.3 (High)
29.5MiniMax-M3
29.2Qwen3.6-Plus
24.4DeepSeek-V4-Flash (Max)
23.7Gemma 4 31B
21.2Step 3.5 Flash (High)
12.7Step 3.7 Flash (High)

推理模型以此图标标识

政务实战能力

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。

政务实战能力（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。

开源模型闭源模型
42.6DeepSeek-V4-Pro (Max)
42.3Claude Opus 4.8 (Max)
42.1MiMo-V2.5
42.0DeepSeek-V4-Pro (High)
41.2DeepSeek-V4-Flash (High)
39.7Qwen3.7-Max
39.5GLM-5.2 (Max)
37.5Kimi K2.6
37.2GLM-5.1
36.8Hy3
36.5Qwen3.7-Plus
35.5Doubao-Seed-2.1-turbo (High)
35.2Kimi K2.7 Code
34.4Gemini 3.5 Flash (High)
33.1MiniMax-M2.7
32.0MiMo-V2.5-Pro
30.8Hy3 Preview
30.4Doubao-Seed-2.1-pro (High)
29.7Grok 4.3 (High)
29.5MiniMax-M3
29.2Qwen3.6-Plus
24.4DeepSeek-V4-Flash (Max)
23.7Gemma 4 31B
21.2Step 3.5 Flash (High)
12.7Step 3.7 Flash (High)

推理模型以此图标标识

Capability

政务行业实战能力分析

从数理审计、复杂业务和合规风控三个任务类型，比较模型在政务场景中的实战表现。

数理审计

评估模型核对数据、检查口径一致性和发现计算异常的能力。

20.0Claude Opus 4.8 (Max)
20.0MiMo-V2.5
20.0Qwen3.7-Plus
20.0Grok 4.3 (High)
0.0DeepSeek-V4-Pro (Max)
0.0DeepSeek-V4-Pro (High)
0.0DeepSeek-V4-Flash (High)
0.0Qwen3.7-Max
0.0GLM-5.2 (Max)
0.0Kimi K2.6
0.0GLM-5.1
0.0Hy3
0.0Doubao-Seed-2.1-turbo (High)
0.0Kimi K2.7 Code
0.0Gemini 3.5 Flash (High)
0.0MiniMax-M2.7
0.0MiMo-V2.5-Pro
0.0Hy3 Preview
0.0Doubao-Seed-2.1-pro (High)
0.0MiniMax-M3
0.0Qwen3.6-Plus
0.0DeepSeek-V4-Flash (Max)
0.0Gemma 4 31B
0.0Step 3.5 Flash (High)
0.0Step 3.7 Flash (High)

Safety & Reliability

政务行业安全与合规

对比模型在政务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。

95.4%Kimi K2.6
95.2%GLM-5.1
93.6%DeepSeek-V4-Flash (High)
92.6%DeepSeek-V4-Pro (High)
92.4%Hy3 Preview
91.8%DeepSeek-V4-Pro (Max)
90.3%Qwen3.7-Max
90.1%Kimi K2.7 Code
89.8%Hy3
89.0%Qwen3.6-Plus
88.5%Grok 4.3 (High)
88.2%Qwen3.7-Plus
88.1%MiniMax-M2.7
87.5%GLM-5.2 (Max)
87.5%MiMo-V2.5-Pro
87.0%MiMo-V2.5
86.0%Claude Opus 4.8 (Max)
85.3%Gemini 3.5 Flash (High)
85.2%MiniMax-M3
80.8%DeepSeek-V4-Flash (Max)
80.3%Gemma 4 31B
77.3%Step 3.5 Flash (High)
75.4%Step 3.7 Flash (High)
71.4%Doubao-Seed-2.1-turbo (High)
67.3%Doubao-Seed-2.1-pro (High)

推理模型以此图标标识

Cost Efficiency

政务行业成本效率

展示模型完成 SQBench 政务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
$0.90DeepSeek-V4-Flash (High)$0.90输入$0.45推理$0.23回答$0.21DeepSeek-V4-Flash (High)
$0.95DeepSeek-V4-Flash (Max)$0.95输入$0.40推理$0.32回答$0.22DeepSeek-V4-Flash (Max)
$1.11MiMo-V2.5$1.11输入$0.56推理$0.33回答$0.22MiMo-V2.5
$1.13Hy3$1.13输入$0.60推理$0.00回答$0.53Hy3
$1.33Gemma 4 31B$1.33输入$1.13推理$0.00回答$0.20Gemma 4 31B
$1.46Step 3.5 Flash (High)$1.46输入$0.91推理$0.00回答$0.55Step 3.5 Flash (High)
$2.24DeepSeek-V4-Pro (High)$2.24输入$0.95推理$0.64回答$0.65DeepSeek-V4-Pro (High)
$2.91MiMo-V2.5-Pro$2.91输入$1.54推理$0.71回答$0.66MiMo-V2.5-Pro
$3.20DeepSeek-V4-Pro (Max)$3.20输入$1.73推理$0.84回答$0.64DeepSeek-V4-Pro (Max)
$3.21Step 3.7 Flash (High)$3.21输入$1.10推理$0.00回答$2.11Step 3.7 Flash (High)
$4.33MiniMax-M2.7$4.33输入$2.27推理$0.49回答$1.58MiniMax-M2.7
$4.40Hy3 Preview$4.40输入$3.53推理$0.00回答$0.87Hy3 Preview
$4.43MiniMax-M3$4.43输入$2.05推理$0.90回答$1.48MiniMax-M3
$4.76Qwen3.7-Plus$4.76输入$2.08推理$1.48回答$1.20Qwen3.7-Plus
$7.07Doubao-Seed-2.1-turbo (High)$7.07输入$2.47推理$1.76回答$2.85Doubao-Seed-2.1-turbo (High)
$9.20Grok 4.3 (High)$9.20输入$6.27推理$2.24回答$0.69Grok 4.3 (High)
$11.25$8.64Qwen3.6-Plus$11.25输入$8.64推理$1.21回答$1.40Qwen3.6-Plus
$11.50GLM-5.1$11.50输入$6.63推理$2.82回答$2.04GLM-5.1
$11.89Kimi K2.7 Code$11.89输入$5.79推理$4.07回答$2.03Kimi K2.7 Code
$15.83Kimi K2.6$15.83输入$7.45推理$3.08回答$5.30Kimi K2.6
$17.44Doubao-Seed-2.1-pro (High)$17.44输入$6.86推理$3.83回答$6.76Doubao-Seed-2.1-pro (High)
$21.13$10.43$7.77GLM-5.2 (Max)$21.13输入$10.43推理$7.77回答$2.92GLM-5.2 (Max)
$24.22$16.27Qwen3.7-Max$24.22输入$16.27推理$4.42回答$3.53Qwen3.7-Max
$54.18$30.95$16.36Gemini 3.5 Flash (High)$54.18输入$30.95推理$16.36回答$6.87Gemini 3.5 Flash (High)
$139.59$46.19$9.87$83.54Claude Opus 4.8 (Max)$139.59输入$46.19推理$9.87回答$83.54Claude Opus 4.8 (Max)

推理模型以此图标标识

Value

政务行业性价比

对比模型政务实战能力与可靠度、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。

政务实战能力 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。

DeepSeekAnthropic小米阿里巴巴智谱月之暗面腾讯字节跳动GoogleMiniMaxxAI阶跃星辰
理想区域政务实战能力可靠度20.030.040.070.0%80.0%90.0%DeepSeek-V4-Pro (Max)DeepSeek-V4-Pro (Max)可靠度91.8%政务实战能力42.6%Click for more infoClaude Opus 4.8 (Max)Claude Opus 4.8 (Max)可靠度86.0%政务实战能力42.3%Click for more infoMiMo-V2.5MiMo-V2.5可靠度87.0%政务实战能力42.1%Click for more infoDeepSeek-V4-Pro (High)DeepSeek-V4-Pro (High)可靠度92.6%政务实战能力42.0%Click for more infoDeepSeek-V4-Flash (High)DeepSeek-V4-Flash (High)可靠度93.6%政务实战能力41.2%Click for more infoQwen3.7-MaxQwen3.7-Max可靠度90.3%政务实战能力39.7%Click for more infoGLM-5.2 (Max)GLM-5.2 (Max)可靠度87.5%政务实战能力39.5%Click for more infoKimi K2.6Kimi K2.6可靠度95.4%政务实战能力37.5%Click for more infoGLM-5.1GLM-5.1可靠度95.2%政务实战能力37.2%Click for more infoHy3Hy3可靠度89.8%政务实战能力36.8%Click for more infoQwen3.7-PlusQwen3.7-Plus可靠度88.2%政务实战能力36.5%Click for more infoDoubao-Seed-2.1-turbo (High)Doubao-Seed-2.1-turbo (High)可靠度71.4%政务实战能力35.5%Click for more infoKimi K2.7 CodeKimi K2.7 Code可靠度90.1%政务实战能力35.2%Click for more infoGemini 3.5 Flash (High)Gemini 3.5 Flash (High)可靠度85.3%政务实战能力34.4%Click for more infoMiniMax-M2.7MiniMax-M2.7可靠度88.1%政务实战能力33.1%Click for more infoMiMo-V2.5-ProMiMo-V2.5-Pro可靠度87.5%政务实战能力32.0%Click for more infoHy3 PreviewHy3 Preview可靠度92.4%政务实战能力30.8%Click for more infoDoubao-Seed-2.1-pro (High)Doubao-Seed-2.1-pro (High)可靠度67.3%政务实战能力30.4%Click for more infoGrok 4.3 (High)Grok 4.3 (High)可靠度88.5%政务实战能力29.7%Click for more infoMiniMax-M3MiniMax-M3可靠度85.2%政务实战能力29.5%Click for more infoQwen3.6-PlusQwen3.6-Plus可靠度89.0%政务实战能力29.2%Click for more infoDeepSeek-V4-Flash (Max)DeepSeek-V4-Flash (Max)可靠度80.8%政务实战能力24.4%Click for more infoGemma 4 31BGemma 4 31B可靠度80.3%政务实战能力23.7%Click for more infoStep 3.5 Flash (High)Step 3.5 Flash (High)可靠度77.3%政务实战能力21.2%Click for more infoStep 3.7 Flash (High)Step 3.7 Flash (High)可靠度75.4%政务实战能力12.7%Click for more info

推理模型以此图标标识