Artificial Analysis Benchmarks DeepSeek Search API: $0.75/1K Searches Delivers +36 Point Score Lift

Background and the Problem

For production autonomous agents, automated research copilots, and real-time knowledge bots, web-search augmented generation (Web-RAG) is the primary safeguard against hallucination and outdated training data. However, deploying search-augmented agents in enterprise applications introduces severe cost bottlenecks: established enterprise search endpoints (e.g., Google, Bing, Perplexity) charge several dollars per thousand queries. In complex multi-turn workflows where recursive search loops execute dozens of queries per task, search overhead quickly eclipses model token expenses. To break this economic bottleneck, DeepSeek rolled out its native DeepSeek Search API, prompting an independent benchmark analysis from Artificial Analysis (@ArtificialAnlys).

Architecture and How It Works

1. High-Throughput Semantic Indexing & Extraction:

• Engineered specifically for LLM token ingestion, filtering boilerplate HTML and tracking tags to return high-density markdown text snippets;

1. Native Turnkey API Ergonomics:

• Integrated directly into the model API payload, eliminating the engineering burden of maintaining bespoke headless scrapers and external rerankers;

1. Disruptive Commercial Economics:

• Priced at an aggressive $0.75 per 1,000 search queries;
• For complex agent tasks requiring an average of ~42 web queries, the total search surcharge amounts to approximately $0.032 per task.

Benchmarks and Measured Results

Artificial Analysis evaluated DeepSeek Search across its comprehensive suite alongside 27 competing endpoints:

• Search Augmented Lift:
• The standalone DeepSeek base model scored 33 points on the Search Index;
• Enabling DeepSeek Search surged the score to 69 points (+36 point increase), ranking 17th of 27 evaluated configurations;
• Sub-Benchmark Breakdown:
• AA-Omniscience (Factual QA): Scored 70% accuracy (10th of 27), proving exceptional retrieval precision on single-hop and multi-hop fact lookups;
• BrowseComp (Complex Multi-step Browsing): Recorded 63% accuracy (25th of 27), indicating room for improvement in deep-link transactional web navigation;
• ROI & Efficiency Leadership:
• Established the definitive benchmark lead for retrieval accuracy per dollar across all tested paid commercial search APIs.

Getting Started for Developers

• API Access: Configurable through standard DeepSeek developer endpoints via the search parameter;
• Architecture Takeaway: Systems deploying high-volume, cost-sensitive agent fleets (such as news summarizers, market monitors, and customer service bots) should leverage DeepSeek Search as an ultra-economical grounding layer, cutting retrieval overhead by up to 80% without sacrificing factual integrity.