DeepSeek Search API Overall first-test score 69 points; ranks lower in complex retrieval
BIBIBI
AT A GLANCE
DeepSeek Search API scored 69 points in the evaluation, performing strongly in fact-finding and offering the lowest search unit price, but performing more weakly in complex retrieval and model processing costs.
Article
DeepSeek First Search Evaluation 69 Score: nearly first in factual Q&A, third from last in complex retrieval
Beating AI Flash: independent evaluation firm Artificial Analysis tested DeepSeek newly launched search API, including it in a comparison of 27 search products. Its overall score was 69, ranking 17. Top-ranked Exa and Perplexity both scored 80.
However,DeepSeek charges only 1000 per 0.75 USD successful searches, making it the cheapest among all paid search API products evaluated.
The evaluation used GPT-5.6 Luna uniformly, having it invoke different search API to complete factual 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 well at directly finding facts, with accuracy reaching 70%, only 3 percentage points behind the leader. But when facing complex questions requiring repeated searches and connecting clues, accuracy was only 63%, ranking third from last among 27 products.
In addition, each search returned an average of about 7000 Token. Although information-rich, this increased the cost for the model to read and process the materials.
Completing one evaluation task cost about 0.032 USD on average, of which search costs were only 0.0085 USD, while model invocation costs reached 0.023 USD, accounting for about 73%of total cost—the second-highest model cost on the list.
Even so,DeepSeek's total cost per question was still the 5 cheapest. It reduced the price of search requests to the lowest level, but lengthy search results consumed part of the cost advantage.
Original link https://m.theblockbeats.info/flash/373160
Key points
01
Artificial Analysis tested DeepSeek's newly launched search API and included it in a comparison of 27 search products.
02
DeepSeek The overall score was 69 points, ranking 17; Exa and Perplexity both scored 80 points.
03
DeepSeek Each 1000 successful search costs 0.75 USD, making it the cheapest among evaluated paid search API products.
04
The evaluation uniformly used GPT-5.6 Luna to invoke different search API; tasks included factual Q&A, deep research, and multi-turn web retrieval.
05
After integrating DeepSeek, the overall score rose from 33 points without internet access to 69 points.
06
DeepSeek Accuracy in directly finding facts was 70%, 3 percentage points behind the leader.
07
Accuracy on complex questions was 63%, ranking third from last among 27 products.
08
Each search returned an average of about 7000 Token。
09
Completing one evaluation task cost about 0.032 USD on average, of which search costs were 0.0085 USD, model invocation costs were 0.023 USD, accounting for about total cost 73%。
10
DeepSeek Total cost per question was the 5 cheapest, while model cost was the second-highest on the list.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This evaluation shows that DeepSeek search API is suitable for directly finding facts, and its search interface itself is inexpensive; however, it ranks lower when multiple searches and connecting clues are required. It returns relatively lengthy content, requiring the model to process more Token, so low interface pricing does not fully translate into lower overall task costs.
Why it matters to readers
Choosing search API should not be based solely on search price; answer quality, complex retrieval capability, returned-content length, and model processing costs must also be compared.
Beginners can understand search API as a tool that provides web materials to models. Low search fees do not necessarily mean the lowest total task cost, because models also incur costs to read and process search results.
Risks and unknowns
The evidence does not include the original Artificial Analysis evaluation report, so the rankings, calculation methodology, and full results cannot be verified from the current materials.
The evidence does not specify the evaluation questions, sample size, or statistical error.
These results are derived from a specific model and task configuration, and it is unclear whether they would still hold after changing the model or use case.
No evidence provided DeepSeek Search API Subsequent real-world usage performance or user feedback.
Related Developments
Loading event timeline…
Related concepts
API
This term is not in the glossary yet. Browse related concepts in the glossary.
