Metaphor
Exa, formerly known as Metaphor, is an AI API designed to connect your AI applications to the internet. Exa is focused on intelligent web search and retrieval. ...
Last verified:
What is Metaphor?
Exa (formerly Metaphor) is a search engine built specifically for AI applications and LLMs. It provides a comprehensive API for search, crawling, and research agents, offering five core functionalities: /search for finding webpages using embeddings-based or keyword search, /contents for obtaining clean parsed HTML from search results, /findsimilar for finding pages similar in meaning to a given link, /answer for getting direct answers to questions, and /research for automating in-depth web research with structured JSON results and citations.
Key features include neural/embedding search technology using transformers for semantic understanding, natural language query processing, instant content retrieval with clean parsed pages, large-scale searches capable of returning thousands of results, content highlights extraction, configurable latency from 180ms to 1s, up to 1,000 results per search, category filtering (news, research papers, companies), domain and date range filtering, and structured outputs optimized for complex queries. The platform leads across FRAMES, Tip-of-Tongue, and Seal0 benchmarks—the most demanding retrieval benchmarks.
Exa is designed for AI developers, coding agents, LLM applications, chatbots, research automation, and agentic workflows. It powers use cases like coding agents searching over GitHub repos and docs, deep research runs, list building, content enrichment, and tracking new events across the web with monitors. Companies like Cognition (co-founder Walden Yan) use Exa to power all parts of Devin, and Cursor uses it for accurate code generation with latest docs.
Metaphor pricing
Pricing model: Freemium
Free Tier: 1,000 requests per month free including web search tool calls, webpage text and highlights, configurable latency 180ms-1s. Search: $7 per 1k requests with up to 10 results (text and highlights included), $1 per 1k requests for each additional result above 10. Deep Search: $12 per 1k requests. Deep-Reasoning Search: $15 per 1k requests. Contents: $1 per 1k pages per content type for full page contents and token-efficient highlights. Monitors: $15 per 1k requests for tracking web events. Answer: $5 per 1k requests. AI page summaries: $1 per 1k pages (billed separately for all endpoints). Agent: $0.025-$2.00 per run with fixed effort modes (Low $0.025, Medium $0.10, High $0.50, X-high $2.00), plus $0.0001 per ACU, $0.007 per search, $0.02 per email, $0.07 per phone number. Enterprise: Custom pricing with volume discounts, SLAs, custom index, and 1:1 support. Startup/Education Grants: $1000 worth of free credits.
Metaphor pros
- Built specifically for AI applications and LLMs, not a wrapper around Google
- Own search engine built from scratch with custom neural search index
- Embedding search technology using transformers for semantic understanding
- Natural language query processing for more accurate results
- Instant content retrieval with clean parsed HTML from any indexed page
- Large-scale searches returning thousands of results for batch processing
- Content highlights extraction for targeted information
- Configurable latency from 180ms to 1s for latency-sensitive agent loops
- Up to 1,000 results per search query
- Best-in-class accuracy on FRAMES, Tip-of-Tongue, and Seal0 benchmarks
- Category filtering for news, research papers, companies, and more
- Domain and date range filtering for targeted research
- Structured outputs with grounded citations for complex queries
- Zero data retention option for enterprise privacy and compliance
- Index updated every hour with constantly added new links
- Startup and education grants offering $1000 free credits
Metaphor cons
- No permanently free tier—only 1,000 requests/month free for evaluation
- Pay-as-you-go pricing may be expensive for high-volume users
- Contents endpoint still billed separately at $1 per 1k pages per content type
- AI page summaries billed separately at $1 per 1k summaries
- Deep-Reasoning Search costs $15 per 1k requests—highest tier
- Monitors endpoint at $15 per 1k requests is expensive
- Contact enrichment for emails ($0.02) and phone numbers ($0.07) billed separately
- Will not crawl pages with noindex tag like any other search engine
- Limited to web search—no support for other data sources
- Beta access required for Agent features
Frequently asked questions about Metaphor
What is Exa?
Exa is a search engine built specifically for AI applications. It was built from scratch to be state of the art at finding high quality information for LLMs. Exa is used by thousands of companies to power their LLM and agentic applications.
What's different about Exa Search compared to traditional search engines?
Traditional search engines are optimized for clicks and ads, and nearly every search API wraps traditional search engines with similar problems. In contrast, Exa is optimized to return the highest quality information for LLM applications. Exa does not make money from ads, so it is fully incentivized to return the highest quality results to customers. Because Exa built its own search engine from scratch, it can provide customized features that other providers cannot.
How is Exa different from LLMs?
Exa is a new search engine built from the ground up. LLMs are models built to predict the next piece of text. Exa predicts specific links on the web given their relevance to a query. LLMs have intelligence and get smarter over time as new models are trained. Exa connects these intelligences to the web.
How can Exa be used in an LLM?
Exa enhances LLMs by supplying high-quality, relevant web content, minimizing hallucination and outdated responses. An LLM can take a user's query, use Exa to find pertinent web content, and generate answers based on reliable, up-to-date information.
How does Exa compare to other search APIs?
Exa.ai offers unique capabilities including: Embedding Search Technology using transformers for semantic understanding, Natural Language Queries for more accurate results, Instant Content Retrieval of clean parsed content for any page in its index, Large-scale searches capable of returning thousands of results for automatic processing, Content Highlights extraction, optimization specifically for AI applications like chatbots and research automation, and Auto search that automatically selects the best search method based on the query.
How often is the index updated?
Exa updates its index every hour and is constantly adding batches of new links. It targets the highest quality web pages. Clients can request specific domains to be more deeply covered—if there is a use-case unlocked by additional domain coverage, they can contact Exa.
How does similarity search work?
When searching using a URL, Exa crawls the URL, parses the main content from the HTML, and searches the index with that parsed content. The model chooses webpages it predicts are talked about in similar ways to the prompt URL, considering factors like text style, domain, and main ideas inside the text. Similarity search is a natural extension for a semantic search engine like Exa and something difficult with traditional search engines.
What security measures does Exa take?
Exa has robust policies and everything is either in standard cloud services or built in house, including their own vector database served in house, own GPU cluster, own query model, and own search solution. Exa can offer unique security arrangements like zero data retention as part of a custom enterprise agreement.
Does Exa have a crawler?
Exa crawls pages on the web, just like any other search engine. If a webpage has the noindex tag and is therefore not crawlable by any search engine, then Exa will not crawl that page.
What's on Exa's roadmap?
Exa's roadmap includes building a (much) larger index and solving search—together with the goal of making search significantly better.