CorporateBench

An agent works through the email archives of four synthetic companies, from a 12-person startup with 353 documents to a 10,000-employee finance firm with 232,692. It rebuilds employees’ reporting and task histories, labels documents by topic, and answers questions that join facts across messages. Answers are graded deterministically against each company’s ground-truth knowledge graph, and every company and every capability carries equal weight in the headline.

Last updated

Main Insights

Results

CorporateBench results, ranked by score (avg@3), highest first. Last updated .
RankModelProviderScore (avg@3)Cost (USD / task)Duration (seconds / task)
1GPT-6.1 SolOpenAI82.2%$0.41437
2GPT-6 AstraOpenAI82.0%$2.42533
3GPT-6 SolOpenAI79.0%$0.58346
4Kimi K3.1 ProMoonshot AI78.4%$0.47461
5Gemini 3.8 FlashGoogle77.6%$1.18593
6Claude Opus 5.5Anthropic75.4%$2.80859
7DeepSeek V4.1 FlashDeepSeek75.3%$0.080709
8DeepSeek V4.1 Flash (1017)DeepSeek75.2%$0.10271
9Claude Opus 5Anthropic74.0%$2.56889
10Muse Spark OpenMeta73.0%$0.45373
11Kimi K3Moonshot AI72.3%$0.61720
12GPT-6 LunaOpenAI71.7%$0.040477