What Exa Actually Is (From Someone Who Builds Research Agents)
I build AI research and sales-intelligence agents for clients, and for the last year Exa has been the search layer in most of them. Exa (formerly Metaphor) is a web search API built for machines, not humans. Instead of matching keywords the way Google does, it embeds both your query and the pages it has indexed into vectors, then finds pages that are semantically close to what you actually asked. That sounds like marketing speak until you run a real query. Ask Google for "startups quietly building AI tools for construction project managers" and you get a list of articles containing those words. Ask Exa the same thing and you get actual startups doing exactly that, because it understood the meaning, not the terms.
That difference is the entire reason to pay for it. Inside an LLM agent, queries are long, natural-language and abstract - "find companies that raised a Series B this year in European climate tech, focused on carbon accounting" - and keyword search falls apart on those. Exa was built for that exact shape of query, and it shows on benchmarks: it scores 81% on the WebWalker multi-hop retrieval test, ten points ahead of Tavily's 71%.
Pricing, plainly. Everything is pay-as-you-go, no subscription, no minimum. Standard search is $7 per 1,000 requests (that is $0.007 a call, contents for the first 10 results included since March 2026). Each extra result above 10 adds $1 per 1,000. Deep Search is $12 per 1,000, Deep-Reasoning $15 per 1,000 - these are the multi-step research modes with structured outputs and citations. The Contents endpoint is $1 per 1,000 pages, AI page summaries another $1 per 1,000. Answer is $5 per 1,000, Monitors (scheduled searches with webhooks) $15 per 1,000. The newer Agent API bills per run: $0.012 at minimal effort up to $1.00 at xhigh, or metered on auto at $0.10 per Agent Compute Unit plus $0.005 per search tool call, with email enrichment at $0.02 and phone at $0.07. New accounts get $20 free plus $10 a month, and startups can apply for $1,000 in credits.
It also ships a free open-source MCP server, which is the single fastest way I have found to wire Exa into Claude, Cursor, VS Code, Windsurf or Gemini CLI in minutes, plus native integrations with LangChain, LlamaIndex and CrewAI.
Why It Pays (The Real Income Angle)
Exa is not a tool you resell directly - it is the engine under products and services. Here are the five ways I have seen solo operators actually bill for it:
1. Research-as-a-service for funds, consultancies and VC firms. This is the cleanest entry point. Institutional research teams hate how Google surfaces their queries and they pay serious money for well-cited, deep research they cannot delegate to an intern in an afternoon. A $500-$2,000 a month retainer for 4-8 sector reports, each 30-50 pages with real citations, is a normal deal. Your tool cost: a few dollars in Exa credits per report at the deep-search rate. The margin is basically the invoice.
2. Sales-intelligence and lead enrichment. The people, companies and papers indexes (1B+, 50M+, 100M+ respectively) make Exa unusually good at "find every startup in Berlin with under 50 employees doing logistics software, then enrich the founders' emails." A clean list of 500 qualified leads sells for $200-$500, and the enrichment API fills in the contacts at $0.02 an email. An agency doing this for five clients a month at $300-$800 each clears a real income with near-zero tool cost.
3. Search-powered agents for small businesses. Local SMBs do not know Exa exists, but they understand "a bot that watches every mention of our competitors and emails us a summary every morning." Build it with Monitors + the Agent API, charge $2K-$5K to build and $200-$500 a month to run. The monitoring cost is trivial - a few hundred monitored searches a month at $15 per 1,000.
4. SEO and market-intelligence products. The webhook-based Monitors endpoint is a ready-made backbone for a productized alert service: competitor pricing changes, new job posts, regulation updates in a niche. Package it at $200-$800 a month per client, run it on $15 per 1,000 requests. Your cost per client is single-digit dollars.
5. Integration consulting. Because the MCP server and the LangChain/LlamaIndex/CrewAI integrations are free, the barrier is low - but most teams still want someone who has already done it. Charging $500-$2K to stand up Exa inside an existing agent stack, with a retainer to maintain it, is real work that takes a day or two per engagement.
The unit economics that sell the story: a client retainer at $500 a month burns maybe $10-40 in API credits at normal search volume. The tool cost is a rounding error inside the invoice.
How I Actually Run It (A Real Workflow)
A typical engagement for me: a VC wants a landscape map of AI-enabled construction software, 25 companies, each with funding, product summary and why-it-matters. I write one long natural-language query per company type - the more specific the better, because Exa rewards specificity - run each through Deep Search, pull the Contents endpoint for the top pages, and feed the whole thing into an LLM with the citations attached. Total Exa spend for the project: about $3-6. Hands-on time: three to four hours. The report bills at $800.
The workflow rule I live by: always run one quick freshness sanity check before believing a result. Exa's index lags on very recent events - it scored 24% on the FreshQA benchmark - so for anything time-sensitive I validate the top hits or pair it with a second source. For evergreen, definitional, who-are-the-players research, it is the best tool I have used, and I have stopped pretending otherwise.
What Exa Is Bad At (Honest)
- It is API-only. No consumer search box, no no-code layer, no dashboard a non-developer can use. If you cannot call a REST endpoint and parse JSON, this tool is not for you, full stop.
- Deep search gets expensive. $12-$15 per 1,000 requests versus $7 for standard search, and research-grade workloads burn hundreds of requests per report. Budget it or you will have a rude credit statement.
- Freshness is a real weakness. Recent-event queries frequently return stale results from the neural index. It scored 24% on FreshQA across 5,000+ queries, so news and pricing use cases need a secondary source.
- Rate limits bite. The default is 10 QPS (600 requests a minute). Multi-agent pipelines need retry and queue logic, and raising the limit means an enterprise sales conversation, not a self-serve plan toggle.
- The index is smaller than Google's. Obscure long-tail pages that exist on the open web are sometimes simply missing. If your niche is tiny, test coverage before committing a client to it.
- Billing discipline is on you. The Agent auto mode meters up to $5 per run, credits are pay-as-you-go, and there is no hard cap unless you build one. A sloppy prompt can quietly cost more than the client invoice covers.
Exa vs The Alternatives
| Tool | Best at | Search style | Free tier | Price |
|---|---|---|---|---|
| Exa AI | Semantic retrieval, research, enrichment | Meaning-based (embeddings) | $20 + $10/mo, ~20K req | $7/1k search, $12-15/1k deep |
| Tavily | LLM-friendly summarized results, easy start | Hybrid keyword + AI answers | Generous free credits | ~$5-8/1k |
| SerpAPI | Mirroring Google's actual rankings | Google SERP wrapper | 100 req/mo | ~$50/1k (varies by engine) |
| Brave Search API | Real-time freshness, low cost | Keyword + web index | 2K req/mo | ~$3-5/1k |
| Perplexity API | Answer-first research | LLM answers + citations | Small credits | ~$5/1k + output tokens |
The honest summary: Tavily is the easy on-ramp, SerpAPI is for SEO work that must match Google, Brave is the cheap freshness play, and Exa wins when your queries are long, natural-language and abstract - which is precisely the case inside an LLM agent.
Who Should Use It (And Who Shouldn't)
Use it if: you build agents, RAG apps or research products and need retrieval that understands meaning; you sell research or sales-intelligence services to funds, consultancies or B2B teams; you want a search layer your LLM can actually reason over; or you are prototyping an agent product and want a generous free tier to prove it works.
Skip it if: you cannot write code; your use case is fundamentally about what happened this week (news, pricing, breaking events) without a second source; you need Google's exact rankings; or your niche is so obscure that you cannot verify Exa's index covers it.
Getting Started (6 Tips)
- Write queries like you are asking a person, not typing keywords. "AI tools used by boutique hotels for revenue management" returns far better results than "hotel AI revenue tool." Specificity is the product.
- Start on the free tier before paying anything. $20 plus $10 a month is a real prototyping budget; run a full test project and only then decide on production spend.
- Use the MCP server for the fastest first demo. Wire Exa into Claude or Cursor in minutes - clients understand "the assistant can search the web properly" faster than they understand an API.
- Cap your Agent runs. Set a fixed effort level (medium at $0.10 is a sane default) instead of leaving auto mode on for long jobs, and add your own spend guardrails.
- Verify freshness for anything time-sensitive. Cross-check recent-event results against a second source; treat Exa as the research engine and something else as the news feed.
- Plan retry and queue logic from day one. The 10 QPS default is real; build for it in the architecture instead of discovering it when the agent starts erroring.
The bottom line: Exa is not a search box you open, it is a search engine you build on. For anyone creating research products, sales-intelligence services or AI agents, it is currently the best semantic retrieval layer in the category - and at $0.007 a search, the engine costs almost nothing next to what the product built on it can bill.