Last updated: 2026-05-21
Exa's neural search API serves 400,000+ developers building AI agents. Free: 1K searches/month. Pro at $7/1K requests. Powers Cursor, HubSpot, and Cognition.
Exa is an AI-native neural search API built for agents and RAG applications, valued at $2.2 billion after its May 2026 Series C. It serves 400,000+ developers at companies including Cursor, HubSpot, and Cognition. Plans start free at 1,000 searches/month; Pro costs $7 per 1,000 requests. Unlike keyword-based alternatives, Exa uses vector embeddings to retrieve semantically relevant web content, returning clean text instead of raw HTML.
Exa is an AI-native web search API built by Exa Labs (San Francisco, founded 2021) specifically for large language models, autonomous agents, and retrieval-augmented generation pipelines. The company was founded by Harvard classmates Will Bryk (CEO) and Jeffrey Wang roughly a year before ChatGPT launched, with the insight that AI would need a search engine designed for machine consumption rather than human browsing. Since launching its API in 2023, Exa has grown to serve more than 400,000 developers and 5,000+ companies, reaching a $2.2 billion valuation on a $250 million Series C funding round led by Andreessen Horowitz in May 2026. The core mechanism is embeddings-based retrieval. Exa encodes every indexed page as a dense vector using its own trained neural networks, then runs nearest-neighbor lookups when a query arrives, surfacing pages whose vectors are mathematically closest to the query. This captures topic similarity, synonymy, and conceptual proximity that keyword APIs miss entirely. In a Fortune 100 enterprise evaluation using the WebWalker multi-hop retrieval benchmark, Exa scored 81% versus Tavily's 71%. Exa offers four search modes: Exa Instant (under 200ms), standard search ($7/1,000 requests), Deep Research ($12/1,000 requests), and Deep Reasoning ($15/1,000 requests). Every request returns clean extracted text instead of raw HTML, reducing token usage in downstream LLM calls. Exa is best for developers building RAG systems, AI coding assistants, autonomous research agents, and sales intelligence platforms. The platform is used in production by Cursor, Cognition, HubSpot, OpenRouter, and Monday.com. Domain-specific indexes cover 1 billion+ people profiles, 50 million+ company records, and 100 million+ research papers. Exa also ships an official MCP (Model Context Protocol) server, free and open-source, that connects Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Gemini CLI, and more than 10 other AI tools to its search capabilities without requiring custom API integration code. Pricing is usage-based with a free entry point. The free tier includes 1,000 searches per month plus $10 in starter credits, with no credit card required. Standard search costs $7 per 1,000 requests; the Agentic tier costs $12 per 1,000 requests. Content extraction is billed separately at $1 per 1,000 requests. The Websets product, which builds structured data collections from search results, starts at $49 per month for 8,000 credits. Enterprise plans include custom indexes, dedicated support, and negotiated rate limits through direct sales. Exa holds SOC 2 Type II certification and the a16z-led Series C funds will scale its systems to handle hundreds of thousands of searches per second.
Free tier: 1,000 searches/month plus $10 in starter credits, no credit card required. Standard search: $7/1,000 requests (raised from $5 in March 2026). Agentic tier: $12/1,000 requests. Deep Research: $12/1,000 requests. Deep Reasoning: $15/1,000 requests. Content extraction billed separately at $1/1,000 requests (full-page retrieval = $8/1,000 total). Websets: $49/month for 8,000 credits. Enterprise: custom via sales. Default rate limit 10 QPS; higher limits require enterprise contract.
Exa is an AI-native web search API built by Exa Labs (San Francisco, founded 2021) that retrieves semantically relevant web content for large language models, autonomous agents, and RAG applications. Unlike traditional search APIs that match keywords, Exa encodes every indexed page as a vector embedding and finds results based on conceptual similarity. The platform serves 400,000+ developers and 5,000+ companies including Cursor, HubSpot, Cognition, and Monday.com. Exa reached a $2.2 billion valuation in May 2026 after raising $250 million in a Series C led by Andreessen Horowitz. The company also ships an open-source MCP server that integrates directly with Claude Desktop, Cursor, VS Code, and other AI tools.
Exa offers a free tier with 1,000 searches per month plus $10 in starter credits, requiring no credit card. Standard search costs $7 per 1,000 requests (raised from $5 in March 2026); the Agentic tier for multi-step workflows costs $12 per 1,000 requests. Content extraction (clean page text) is billed separately at $1 per 1,000 requests, bringing full-page retrieval to $8 per 1,000 total. The Websets product for building structured data collections starts at $49 per month for 8,000 credits. Enterprise pricing is custom via sales and includes higher rate limits above the default 10 QPS cap.
Exa's core features include neural semantic search using vector embeddings, which scored 81% on the WebWalker multi-hop retrieval benchmark. Exa Instant delivers results in under 180 milliseconds, suitable for real-time agent applications. Domain-specific indexes cover 1 billion+ people profiles, 50 million+ company records, and 100 million+ research papers. The query-dependent highlights feature extracts only the relevant passages from each page, fitting 4-5 more sources into the same LLM token budget. An official MCP server connects Exa to Claude, Cursor, VS Code, and Windsurf at no additional cost.
Yes, Exa has a free tier that includes 1,000 searches per month plus $10 in starter API credits, with no credit card required. The free tier provides access to neural search, keyword search, and content extraction features. Beyond the free tier, pricing is usage-based: standard search costs $7 per 1,000 requests and content extraction adds $1 per 1,000 requests. Developers can also access the Exa MCP server at no cost to connect Claude, Cursor, and other AI tools to Exa's search capabilities.
The closest competitor is Tavily, an agent-first search API that scored 71% on WebWalker multi-hop retrieval (versus Exa's 81%), with tighter out-of-the-box integration with LangChain and LlamaIndex. Serper provides parsed Google SERP data at $0.30 per 1,000 queries, making it the cheapest option for high-volume keyword search without semantic capabilities. Brave Search API runs on an independent index of 35 billion+ pages and is the strongest choice for privacy-sensitive applications in healthcare or legal. Perplexity Sonar returns pre-synthesized cited answers in a single API call, which suits use cases where the LLM reasoning should happen inside the search layer rather than in a custom agent.
Exa is best for ML engineers and developers building retrieval-augmented generation (RAG) pipelines who need semantically relevant results rather than keyword matches. Teams at AI-native companies building coding assistants, research agents, and sales intelligence platforms benefit most from Exa's neural index and content extraction. Companies in the Cursor, Cognition, and HubSpot tier (Series A and above with production agent workloads) form the primary customer base. Exa is not suitable for non-technical users without API integration experience, or for teams that need historical web archive data, as Exa indexes only the live web.
The API is the core product. Exa provides a REST API and official SDKs for Python (exa-py) and JavaScript/TypeScript (exa-js). The API supports neural search, keyword search, hybrid search, content extraction (clean page text), highlights (relevant passages), and similarity search (finding pages similar to a given URL). Exa also ships a free, open-source MCP (Model Context Protocol) server that exposes web search, company research, LinkedIn search, and deep researcher capabilities to Claude Desktop, Cursor, VS Code, Windsurf, Gemini CLI, and other AI tools without requiring custom API code. The default rate limit is 10 QPS, with higher limits available via enterprise contracts.
On the WebWalker multi-hop retrieval benchmark, Exa scored 81% versus Tavily's 71%, a 10-point gap that compounds across complex agent workflows requiring multiple hops of evidence. For standard search, Exa charges $7 per 1,000 requests versus Tavily's pay-as-you-go rate of $0.008 per credit. Tavily offers stronger out-of-the-box integration with LangChain and LlamaIndex, while Exa's MCP server gives it an edge for Claude, Cursor, and VS Code users. Tavily was acquired by Nebius (an AI cloud company) in February 2026, while Exa raised its own independent $250M Series C at a $2.2B valuation in May 2026. For semantic research tasks and multi-hop workflows, Exa is the stronger choice; for developers already in the LangChain ecosystem or with tight per-query cost constraints, Tavily is a reasonable alternative.