Brave Search on Friday unveiled a revamped API designed to significantly enhance the utility of web search for artificial intelligence applications, alongside research demonstrating that less powerful, open-weight Large Language Models (LLMs) can outperform industry leaders like ChatGPT, Google AI Mode, and Perplexity when leveraging the API’s higher-quality data.
The updates center around the newly released LLM Context API, which Brave claims is the most powerful search API to date, optimized for providing LLMs with relevant web context. The company is already utilizing the API internally, powering over 22 million answers daily within Brave Search, making it the largest private, user-facing AI application globally. Internal testing, conducted using Claude Opus 4.5 and Claude Sonnet 4.5 as judges, showed that Inquire Brave – powered by Brave’s LLM Context and the open-weights Qwen3 model – outperformed competitors in head-to-head comparisons.
“To date, the AI industry has emphasized the importance and value of high-finish models, but our testing shows that less powerful open-weights models can outperform closed frontier models if they incorporate high-quality grounding data,” Brave stated in a press release. The LLM Context API provides this data, ranking and compiling relevant information in a format optimized for LLM consumption. This is made possible, Brave asserts, by its complete search engine infrastructure, a key differentiator from data scraping methods.
The evaluation, conducted on November 30, 2025, involved collecting answers from Ask Brave, Grok, Google AI Mode, ChatGPT, and Perplexity using a set of 1,500 queries sampled from real-world user searches, such as “does iphone collect data when off.” Answers were then evaluated by LLMs-as-judges, considering all pairwise comparisons to maximize reliability and controlling for position bias.
The results, as presented by Brave, show Grok achieving the highest average rating (4.71), closely followed by Ask Brave (4.66). Google AI Mode (4.39) and ChatGPT (4.32) trailed behind, with Perplexity receiving the lowest average rating (4.01). Notably, Ask Brave’s performance was achieved using the Qwen3 model, which Brave characterizes as a lower-performance open-weights model. The key to its success, according to the company, was the superior grounding context provided by the LLM Context API.
Beyond the LLM Context API, Brave is also releasing Brave Search API Skills and an integrated API assistant within its Developer Portal. These tools are designed to simplify the development process and provide support for developers utilizing the API. Skills are also accessible on developer AI tools like Cursor, OpenCode, and ClaudeCode.
Brave is also introducing two recent API plans: Search and Answers. The Search plan, priced at $5 per 1,000 requests, includes access to Web, LLM Context, Images, News, Videos, and other search types. The Answers plan provides researched responses to questions, alongside web results, and is priced at $4 per 1,000 web searches plus $5 per million tokens (input and output). Both plans include $5 of free credit each month.
Brave emphasizes its position as one of only three independent, global-scale search indexes in the western world outside of Huge Tech, and the only one offering an open API with options for SOC2 compliance and Zero Data Retention. The company highlighted the benefits of avoiding the complications associated with web scraping, including potential Terms of Service violations and data access limitations. Brave also stated it does not use search queries to train its own LLMs and offers Zero Data Retention, ensuring no queries are stored or logged.
The LLM Context API functions by performing a standard web search on Brave’s index, then extracting and ranking “smart chunks” of data from relevant pages. This process includes clean text extraction, structured data extraction, specialized code context extraction, forum discussion extraction, and YouTube caption handling. The API also supports Goggles, Brave’s feature for filtering and boosting search results, and an LLM Context budget for controlling token usage. Localized context can be incorporated by passing user location via headers.
