律动BlockBeats|Oct 22, 2026 06:59
[DeepSeek Search First Test Score: 69 Points – Near Top in Factual Q&A, Third from Last in Complex Retrieval]
Beating AI Newsflash: Independent evaluation agency Artificial Analysis tested the newly opened DeepSeek search API, comparing it against 27 search products. DeepSeek achieved a composite score of 69, ranking 17th overall. The top spots were taken by Exa and Perplexity, both scoring 80 points. However, DeepSeek charges only $0.75 per 1,000 successful searches, making it the cheapest paid search API among all participants.
The evaluation uniformly utilized GPT-5.6 Luna, which called different search APIs to complete tasks involving factual Q&A, in-depth research, and multi-round web retrieval. After integrating DeepSeek, the composite score improved from 33 (offline) to 69. DeepSeek performed well in direct factual searches, achieving a 70% accuracy rate, just 3 percentage points behind the top scorer. However, when tackling complex problems requiring iterative searches and clue assembly, its accuracy dropped to 63%, ranking third from last among the 27 products.
Additionally, each search on average returned approximately 7,000 tokens. While the information was abundant, it increased the cost of model reading and processing. Completing a single evaluation task cost an average of $0.032, with search expenses accounting for only $0.0085, while model invocation costs reached $0.023—about 73% of the total cost, the second-highest model cost on the leaderboard. Despite this, DeepSeek's total cost per task remained the fifth cheapest overall. It minimized search request prices to the lowest level, but the lengthy search results consumed part of its cost advantage. [Original Link]
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink