Exa
AI-native neural web search API for agents
The semantic search API that finds pages keyword engines miss, worth it when your agent needs meaning over matches and someone is watching the meter.
Dev teams building RAG and research agents who need neural discovery, find-similar, and clean LLM-ready content in one call.
Cost stacks fast at volume. Per-request pricing plus add-ons for extra results, summaries, and Agent effort means the bill outpaces request count.
Reach for it when semantic recall is the job and engineering owns the caching. For high-volume lookup on a fixed budget, a cheaper keyword API wins.
What you're actually getting
Exa is a neural search API that ranks pages by meaning instead of keywords. It crawls the web and builds its own embeddings index, so results come straight from that crawl with no keyword engine in the middle.
The core move is find-similar and semantic discovery. Ask for "companies building RAG for legal docs" and it returns actual homepages, where a keyword engine hands back listicles that happen to match the words.
Beyond search it ships Contents, Answer, Agent, Monitors, and Websets, plus vertical indexes for people, companies, and code. It is a toolkit for getting the web into your model, built for agents that read content.
exa.ai returns the page content with freshness controls built in so "what happened this week" doesn't quietly hand back something from last month (it's search is built for a model to read, so it wasn't too hard to train into the model's weighted distribution)
Where it earns its keep
It shines on conceptual queries keyword search misses. Obscure repos, niche blog posts, and "pages like this one" are where the neural index pays off.
The Highlights feature pulls the 2-3 relevant passages per result instead of dumping full pages, which cuts token usage hard when your agent fires search on every turn.
Real customers back the quality. Cursor, HubSpot, and Cognition (Devin) run it in production, and third-party benchmarks put Exa ahead on FRAMES and Tip-of-Tongue retrieval.
Exa provides an agent-friendly web search experience that nicely balance latency and breadth of search results.
Where it'll bite you
Cost is the first trap. Base search is $7/1k, but extra results past ten, AI summaries, per-content-type Contents billing, and Agent effort modes all stack on top. One Reddit CTO logged $4,800/month on a 22-client enrichment pipeline and said the cost was eating his margins.
Freshness is the second. The neural index does not update as fast as Google's live crawl, so "what happened yesterday" queries land weaker. Builders report pairing a news API with Exa for recency.
Exact-string lookups suffer too. Neural retrieval is the wrong tool for error messages and verbatim snippets. Add caching, dedup, and a rate-limit plan from day one, because the 10 QPS default ceiling catches parallel agent pipelines mid-demo.
We built it on Exa and the execution is very good... That's roughly $1,200 a week, so $4,800 a month just for search infrastructure. The output quality is pretty good, the briefs are being used by the sales teams and we've seen a measurable uptick in conversion, so the product works. The problem is that the infrastructure cost starts eating into the margin of the service itself.
What it costs, really
There is no subscription. You load prepaid credits and draw down per request. New accounts get $20 in credits plus $10/month free, enough to prototype.
Standard search runs $7 per 1,000 requests, Deep Search $12 to $15, Contents $1 per 1,000 pages per content type, and the Agent endpoint $0.012 to $1.00 per run depending on effort. Enterprise is a custom quote with volume discounts and SLAs.
The sticker price is the small part. The real bill rides on add-ons, result count, and Agent effort, so price your own pilot at production volume before trusting anyone else's number.
Same job, other tools
Also filed under Web data & extraction infrastructure:
