Parallel Search has launched a web search API specifically designed for AI agents, leveraging a proprietary web index. Unlike traditional search engines that cater to human users by ranking URLs, this API focuses on providing relevant data in a token-efficient manner. This means it optimizes for what tokens should be included in an AI’s context window, enhancing the agent's ability to reason and complete tasks effectively. The architecture includes semantic objectives that capture intent, token-relevance ranking, and information-dense excerpts, all intended to streamline AI search. The performance of Parallel Search stands out in benchmarks, especially for complex queries that require multi-hop reasoning. In tests involving multi-hop tasks and simple queries, Parallel outperformed other search APIs, achieving higher accuracy while also being more cost-effective. For instance, in the HLE benchmark, Parallel achieved 47% accuracy at a cost of $82 per 1,000 queries, while competitors like Exa and Tavily had lower accuracy and higher costs. Similarly, in the BrowseComp benchmark, Parallel reached 58% accuracy at a cost of $156, significantly surpassing the performance of others. Testing was conducted with a sample of 100 questions from a larger set of 2,500, with evaluation methods relying on OpenAI’s models. The benchmarks assessed various aspects of AI search capabilities, such as web traversal and reasoning across multiple sources. With these results, Parallel Search addresses a growing need for sophisticated search solutions tailored for AI, moving beyond the limitations of traditional search methodologies.