Brave has launched AI Grounding with the Brave Search API, a comprehensive solution that connects AI outputs to verifiable web data, improving search performance in AI applications. This technology anchors Large Language Model (LLM) responses in high-quality, factual information from reliable sources, reducing hallucinations and enhancing nuanced responses.
Key Performance and Usage
AI Grounding achieves state-of-the-art (SOTA) results on the SimpleQA benchmark with an F1-score of 94.1%, despite not being specifically optimized for this test. It currently powers the "Answer with AI" feature on Brave Search, which processes over 15 million queries daily, and will soon support advanced Research-mode features in Brave Search and Brave Leo, Brave’s integrated AI assistant. Brave Search API also supplies real-time web search data to most of the top 10 AI LLMs, sometimes as the sole search engine index provider.
Access and Pricing
Customers can access the AI Grounding endpoint via new public AI plans:
- AI Grounding plan (new): Grounded answers using single or Research mode (multiple searches, reasoning, planning). Priced at $4 per 1,000 web searches plus $5 per million tokens (input and output).
- Pro AI plan (update): Existing openapi-compliant endpoint access with unchanged pricing.
Research mode is available now for API customers and will soon enhance deep research features for Brave Search and Brave Leo users.
Quality Assessment and Benchmarking
Brave evaluates quality through AI-based and human assessments, with a dedicated team ensuring consistency. The SimpleQA benchmark, developed by OpenAI, tests factual accuracy on short, fact-seeking questions with indisputable answers across diverse domains. Although designed for LLMs without external data, it also measures grounding capabilities when LLMs use web context.
Brave compares favorably to industry peers on SimpleQA, with a detailed analysis forthcoming to interpret benchmark results and limitations.
Single-Search vs. Multi-Search Approaches
A critical factor in SimpleQA performance is the number of search queries per question:
- Single-search systems issue one web query, then generate an answer. This is fast and cost-efficient, with Brave Search responses streaming in under 4.5 seconds on average.
- Multi-search systems perform sequential searches, refining queries iteratively. This increases API calls, context size, reasoning time, and compute cost, often extending response times to minutes. Such systems suit background or high-accuracy tasks but are less ideal for real-time use.
Brave Search API delivers SOTA performance in both single-search and multi-search modes, balancing speed, cost, and accuracy depending on application needs.
