Quick Verdict
Exa is an independent web search and content-retrieval platform built for AI applications. Metaphor was its former brand, so old Metaphor pricing and endpoints should not be treated as current. The platform now spans Search for web results and optional contents, Contents for known URLs, Deep Search for complex queries, Agent for asynchronous research and structured tasks, and Monitors for recurring updates. It fits agents, RAG, coding assistants, and research products rather than ordinary users looking for a free search box.
Adoption requires testing retrieval quality, latency, contents, data terms, and the combined bill. Exa’s standard privacy policy says Query Data is used to improve products and technology, including training and fine-tuning models, and warns users not to submit personal information in open query fields. Zero Data Retention is not presented as a default self-serve entitlement. It is an Enterprise arrangement. Do not assume ZDR for confidential or regulated workloads without the applicable enterprise agreement. Alternatives include Tavily, Firecrawl, Brave Search API, and Perplexity.
Best For
- AI agent and RAG engineers who need search, contents, and citations through APIs.
- Coding-assistant teams retrieving current repositories, docs, and changelogs.
- Research and market-intelligence products using deep or structured workflows.
- Teams that can evaluate retrieval quality and control budgets, keys, caching, and retries.
- Enterprise buyers prepared to contract separately for ZDR, HIPAA mode, SLAs, or custom security.
It is a poor fit for casual search users or anyone planning to put personal information, trade secrets, or regulated queries into a standard self-serve account. Contractual processing terms for business customers must be confirmed rather than inferred from the public API.
Key Features
- Search API: Returns real-time results, optional page contents, and highlights, with configurable latency from roughly 180ms to one second and up to ten results in the base request.
- Semantic and auto retrieval: Supports natural-language meaning, URL-based similarity, and an auto mode that chooses a retrieval method for the query.
- Contents API: Retrieves full text, highlights, summaries, links, or subpages for known URLs as LLM-ready context.
- Deep Search and Answer: Runs deeper retrieval for complex questions or generates search-grounded answers with citations; source verification remains necessary.
- Agent and structured output: Handles asynchronous research, list building, and enrichment with schema-shaped, cited results.
- Monitors: Repeats searches on a schedule and delivers new events through webhooks.
- Developer tooling: Offers Python and TypeScript SDKs, an MCP server, and integrations with agent and RAG frameworks.
- Independent index: Exa says it built its search engine from scratch and updates the index hourly, although depth still varies by domain, region, and language.
Use Cases
- Use low-latency Search inside agent tool loops, then request full text or highlights only when needed.
- Feed fresh URLs, dates, and source text into RAG and fact-checking pipelines.
- Retrieve repositories, documentation, Stack Overflow, and changelogs for coding agents.
- Run multi-step research and structured list building through Deep Search or Agent.
- Monitor companies, products, policies, and news on a schedule with webhooks.
- Expand from a known URL into semantically and stylistically related pages.
Pricing
As of July 21, 2026, the site lists $20 in sign-up credits and $10 in monthly Free Tier credits. Search costs $7 per 1,000 requests with up to ten results per request; each additional result above ten costs $1 per 1,000 requests. Contents costs $1 per 1,000 pages per content type, and AI page summaries cost another $1 per 1,000 pages. Deep Search costs $12 per 1,000 requests, while Deep-Reasoning Search and Monitors cost $15; Answer costs $5.
Agent fixed-effort modes range from $0.012 to $1 per run. The default auto effort varies compute and tool use, while searches and contact enrichment are additional usage components. Enterprise uses custom pricing for higher volume, custom indexes, SLAs, MSAs, support, and ZDR. Total cost can combine result count, Contents types, summaries, Agent tool calls, partner data, and retries.
| Endpoint | Public base price at cutoff | Main boundary |
|---|---|---|
| Free Tier | $20 sign-up + $10 monthly credits | Credits, not unlimited requests |
| Search | $7/1,000 requests | Up to ten results; extras cost more |
| Contents | $1/1,000 pages/content type | Text, highlights, or summaries can add components |
| Deep / Deep-Reasoning | $12 / $15 per 1,000 | More complex, multi-step retrieval |
| Agent | $0.012-$1 per fixed-effort run | Auto effort, search, and enrichment can add cost |
| Enterprise | Custom | Volume, support, custom security, and ZDR |
Pros
- Search, contents, highlights, summaries, and structured output are designed around LLM context.
- An independent index and semantic retrieval can surface long-tail pages missed by keyword matching.
- Search, Deep, Agent, and Monitors cover both low-latency tool calls and asynchronous research.
- Python, TypeScript, MCP, and common agent/RAG integrations are available.
- Enterprise materials state SOC 2 Type II and provide a path to DPA, ZDR, and custom security terms.
Cons
- Standard Query Data may be used for training and fine-tuning, making raw sensitive queries inappropriate.
- ZDR is an Enterprise arrangement, not a public default promise for free or pay-as-you-go accounts.
- Result count, content types, summaries, Agent, and enrichment make billing more complex than one search price.
- Regional, non-English, dynamic-page, and domain-specific coverage require production-query testing.
- Search and generated results can omit, age, or mismatch information; citations still need review.
- SOC 2 does not automatically satisfy every industry’s compliance, residency, or retention requirements.
Alternatives
| Tool | Better for | Advantage | Tradeoff |
|---|---|---|---|
| Tavily | Agent-native search, extract, crawl, and research APIs | Direct agent workflow and mature integrations | Retrieval style, prices, and data terms need separate comparison |
| Firecrawl | Known-site crawling and Markdown extraction | More focused crawling and site traversal | Discovery and an independent search index are not its only focus |
| Brave Search API | Independent conventional search and SERP data | Clear search-index boundary | LLM content and semantic workflows need more assembly |
| Perplexity | Human-facing research and answers | Mature answer and citation interface | Not positioned as the same retrieval infrastructure layer |
FAQ
What is the relationship between Metaphor and Exa?
Metaphor was Exa’s former brand. Current product, documentation, pricing, and contracts should use Exa; the old Metaphor page is only a migration source.
How much free usage does Exa provide?
At the cutoff, the pricing page lists $20 in sign-up credits and $10 in monthly Free Tier credits. Limits and prices can change, so the dashboard remains authoritative.
Does Exa train on query data?
The standard privacy policy says Query Data is used to improve products and technology, including training and fine-tuning models, and tells users not to submit personal information in query fields.
Does Exa provide Zero Data Retention by default?
It should not be assumed. Exa lists ZDR as an Enterprise capability requiring a custom arrangement; no public default ZDR promise is stated for standard self-serve accounts.
What security assurance does Exa publish?
Exa states that it is SOC 2 Type II certified and provides SOC 2 reports, a DPA, and other documentation through its Trust Center. It also publishes a vulnerability-disclosure policy. Buyers still need to validate the applicable plan and contract.
Can Search and Contents create separate charges?
Yes. The Search request has a base result allowance, while contents, content types, summaries, and additional results can add charges. Validate real parameters against dashboard billing rather than budgeting from $7 per 1,000 alone.
What should a team test before adopting Exa?
Use production queries to measure recall, source quality, extraction completeness, freshness, language coverage, P95 latency, and total task cost. Review query training, retention, and enterprise security terms at the same time.
Bottom Line
Exa has grown beyond the old Metaphor semantic-search API into a platform combining Search, Contents, Deep, Agent, and Monitors. Its independent index, LLM-oriented output, and retrieval layers are useful infrastructure. The tradeoffs are compound billing, regional-quality testing, and standard Query Data use for model training and fine-tuning.
Start with free credits and a representative evaluation set, then budget the entire call chain. Do not send personal, confidential, or regulated queries through standard open query fields. If ZDR, HIPAA mode, SLAs, or custom security are mandatory, sign the appropriate Enterprise agreement and review the Trust Center evidence before production launch.