Overview
Exa is a search API designed for AI agents, offering search, crawling, and research capabilities through one API. It delivers high-accuracy, token-efficient results with highlights, structured JSON output, and low latency.
Our take
Exa is web search designed to be consumed by AI agents rather than by people, with a generous free allowance and per-request pricing after it. That is a real distinction: results are structured for programmatic use and retrieval quality is tuned for feeding a model rather than for a human scanning a results page.
Worth it if
Developers building retrieval-augmented applications and agents that need current web content.
Look elsewhere if
Human research — a conventional search engine is faster if you are the one reading.
✨ Features
Agent Search
High-accuracy web search purpose-built for AI agents.
Token Efficiency
Highlights extract relevant excerpts, cutting tokens by up to 90%.
Structured Output
Returns extracted web data as structured JSON.
Research Agent
Async deep research and content crawling endpoints.
⚖️ Pros & Cons
Pros
Cons
💰 Pricing
Free
$0 /mo
- 1,000 search requests/month
- Web search tool calls for agents
- Webpage text and highlights
- Configurable latency options
Search
$7 /1K requests
- Real-time web search data
- Webpage text and highlights
- Agent integration
- Contents from $1/1K pages
Deep Search
$12 /1K requests
- Multi-step agent workflows
- Structured outputs
- Web-grounded citations
- $12-15 per 1K requests
Enterprise
Custom
- Volume discounts
- Higher rate limits
- Dedicated support
- $1,000 startup/education grants
Categories
❓ Frequently Asked Questions
Why use this instead of a general search API?
Because the output shape differs. Search built for agents returns cleaned content and semantically relevant passages rather than ranked links and snippets designed to attract clicks. For a model consuming the result, that difference reduces token spend and improves answer quality more than raw ranking quality does.
How does per-request pricing behave in an agent?
It multiplies, because agents search repeatedly within a single task. One user question can generate several searches plus content fetches. Instrument searches per completed task before scaling, and cap them — an agent looping on searches is both the most common cost surprise and usually a sign of a prompt problem.