DeepSeek Search's first test scores 69 points: factual Q&A close to first, complex retrieval third from last.

Beating AI News Flash: Independent evaluation agency Artificial Analysis tested DeepSeek's newly opened search API, including it in a comparison of 27 search products. It received an overall score of 69, ranking 17th. The top-ranked Exa and Perplexity both scored 80.

However, DeepSeek charges only $0.75 per 1,000 successful searches, making it the cheapest among all paid search APIs tested.

The evaluation uniformly used GPT-5.6 Luna, having it call different search APIs to complete fact-based Q&A, deep research, and multi-turn web retrieval. After integrating DeepSeek, the overall score rose from 33 without internet access to 69.

DeepSeek performed relatively well in direct fact-finding, achieving a 70% accuracy rate, only 3 percentage points behind the top rank. But when facing complex questions requiring repeated searches and stitching together clues, its accuracy was only 63%, ranking third from the bottom among the 27 products.

In addition, it returns an average of about 7,000 tokens per search. Although information-rich, this increases the cost of the model reading and processing the materials.

The average cost to actually complete one evaluation task was about $0.032, of which the search fee was only $0.0085, while the model invocation cost reached $0.023, accounting for about 73% of the total cost—the second-highest model cost on the list.

Even so, DeepSeek's total cost per question remains the 5th cheapest. It has pushed the price of search requests to the lowest level, but lengthy search results eat away part of its cost advantage.