Exa AI

Exa is a neural web search engine built for AI agents and developers, not humans. Instead of matching keywords like Google, Exa uses embeddings to understand the meaning of a query and return semantically similar pages. The result is far cleaner retrieval for RAG pipelines, research agents, and LLM tools. Exa is API-first — there's no consumer search box like Google or Perplexity — so it's designed to be called from code rather than used in a browser.

Exa gives every developer 1,000 free search requests per month. Beyond that, pricing is pay-per-request: standard neural search with contents is $7 per 1,000 requests, Exa Instant (sub-200ms) is $5 per 1,000, deep search is $12 per 1,000, and deep-reasoning is $15 per 1,000. Contents and AI summaries are $1 per 1,000 pages. There are no seats, no monthly minimums, and enterprise plans with custom rate limits are available on request.

Exa Instant is a sub-200ms neural search mode launched in February 2026, designed specifically for real-time agentic workflows where latency kills UX. It trades a little depth for dramatic speed gains — perfect for chat copilots, voice agents, and any LLM loop where users are waiting for a response. At $5 per 1,000 requests it is also Exa's cheapest tier, making it the default choice for high-volume agent use cases.

For pure semantic retrieval, most developers rate Exa as the strongest of the three. Tavily is optimized for LLM-friendly summarized results and bundles AI answers inline. SerpAPI is a wrapper around Google SERPs — it's great if you specifically need Google's ranking, but it's keyword-based. Exa wins when your queries are long, natural-language, or abstract concepts that don't have obvious keywords, which is the common case inside an LLM agent.

Yes. Every Exa account gets 1,000 free search requests per month — no credit card required. That's enough to prototype a full RAG application or build a working research agent before you pay anything. Startups and education projects can also apply for $1,000 in free credits, which extends the free runway substantially. Once you're in production you only pay for what you actually use.

Exa does expose a web interface at exa.ai for quick semantic search demos, and Exa Websets lets you build curated collections of pages without code. However, Exa's real product is the API — if you just want a consumer AI search engine, Perplexity, Andi, or You.com will feel more natural. Exa is best when you're building something on top of search.

Common agent use cases include grounding LLM answers in fresh web data, finding competitor pages by description, curating training datasets for fine-tuning, powering research assistants that read and cite sources, running market intelligence workflows, and building internal Perplexity-style search over arbitrary domains. Exa is wired into LangChain, LlamaIndex, CrewAI, and the Claude and OpenAI tool-use ecosystems, so integration is usually a few lines of code.

Exa is a San Francisco startup that raised an $85M Series B in 2026, bringing total funding to over $100M. It's backed by Lightspeed, Nvidia, Y Combinator, and a number of prominent AI founders. The company has grown quickly by positioning itself as the search engine for AIs rather than for humans — a deliberately niche bet on the assumption that most web search traffic will soon come from agents, not people.