Exa

What is Exa?

How does Exa work?

Use Cases

Pricing

Pricing model: Freemium

Common Limitations

• 10 QPS on Search is a hard ceiling for concurrent agent workloads · Exa's default rate limit is 10 queries per second on /search and /answer, and 5 QPS on Deep Search. This limit applies to the team as a whole across all API keys, so adding keys does not raise it. The upgrade to 25 QPS is automatic only after purchasing $1,000 in credits within a 30-day window—spending $1,000 does not qualify, only buying it does—and the increase lasts 90 days before requalification. For a single agent making sequential calls, 10 QPS is ample. For a fleet of agents running in parallel, or for a batch enrichment job, it is a burst bottleneck that produces HTTP 429 responses and requires exponential backoff on the client side. A published GAIA validation campaign recorded 1,341 HTTP 429 errors from Exa, accounting for 44% of all tool failures in that run—the single largest source of failure by a wide margin[reference:0]. That is not a benchmark of Exa's search quality; it is a benchmark of what happens when an agent fleet meets a 10 QPS limit without adequate retry logic.
• Statistical parity with competitors, not dominance, on agent search benchmarks · Exa's own marketing emphasizes semantic accuracy, and on entity-lookup benchmarks it does outperform Tavily and Firecrawl substantially—75.5% top-result precision on people profiles versus 40.5% and 41.6%, and 81.5% on company lookups versus 61.3% and 61.7%[reference:1]. But those benchmarks are run by Exa. On the independent AIMultiple agentic search benchmark of 100 AI/LLM queries across eight APIs, Exa scored 14.39, statistically indistinguishable from Brave (14.89), Firecrawl (14.58), and Parallel Search Pro (14.21) because the confidence intervals overlap[reference:2]. In the Artificial Analysis Search Index for October 2026, Exa auto scored 74, one point behind Parallel Search advanced (75) and six behind Perplexity Search (80)[reference:3]. The honest read is that Exa is a strong default for semantic and entity search, not a universal leader across every query type and every benchmark. Teams that need the absolute highest score on a specific benchmark should run their own evaluation rather than assuming Exa's entity-lookup advantage generalizes.
• The index is the boundary: recency is bounded by crawl cadence · Exa runs semantic search over its own index, built from a crawl of more than 500 billion webpages. That architecture is the source of its accuracy on entity lookups and its ability to return page text in the same response. It is also the source of its hardest limitation: recency is bounded by crawl cadence, not by live retrieval. A page published minutes ago may not be in the index. Independent analysis notes that “very direct factual lookups can return semantically-similar-but-wrong matches” when the index does not contain the exact target[reference:4]. Exa's own alternatives guide concedes that competitor comparisons sometimes point to “index freshness for brand-new content” as a drawback[reference:5]. For workflows that require the most recent version of a page, or that need to read a specific URL's live content, Exa's Contents endpoint returns what the crawler last captured—not a fresh fetch. Firecrawl renders JavaScript on a live fetch; Exa returns the cached snapshot. The two are often used together for exactly this reason: Exa finds the pages, Firecrawl fetches them live when freshness matters.
• Search does not paginate, and category filters have documented gaps · Exa's search endpoint does not support pagination. numResults accepts 1 to 100 with a default of 10, and the API returns that many results and stops. There is no cursor or offset for retrieving the next page. For an agent that needs to exhaustively enumerate results—competitive research, list building, due diligence across many sources—the only option is to issue multiple queries with different phrasings and deduplicate client-side. Separately, the company and people categories do not support startPublishedDate, endPublishedDate, or excludeDomains filters; using them returns HTTP 400. The lite search type serves cached content only and rejects contents.maxAgeHours with any value, including -1[reference:7]. These are documented spec constraints, not bugs, but they are not surfaced prominently during evaluation, and they force workflow redesign when discovered in production.
• Latency variance makes sub-second guarantees unreliable · Exa's median latency in vendor-run benchmarks is attractive—472 ms in one comparison run, against 1.3 s for Perplexity Sonar and 3.6 s for Tavily[reference:8]. But independent latency measurements show substantial variance: Auto latency on non-news queries was 1.96 s p50 on October 13 and 6.6 s on October 14, with server time pinned at almost exactly 6.05 s on every query in the second measurement[reference:9]. That pattern—a bimodal distribution with a long tail—means an agent designed around a 500 ms budget will occasionally see 6-second responses. For real-time applications such as customer-facing assistants or autocomplete, the variance is more consequential than the median. The Content endpoint is faster (100 QPS) and returns cached text in milliseconds when the page is stored, but search itself is not a low-latency guarantee under load.
• No end-user app, and compliance features are Enterprise-only · Exa is developer infrastructure. There is no end-user research interface, no dashboard for non-technical users, and no built-in UI beyond the billing and API key management pages. Zero Data Retention, HIPAA, and SSO are reserved for Enterprise plans; Free and pay-as-you-go users do not have these controls[reference:10]. For teams evaluating Exa as part of a regulated workflow, this means the self-serve tier cannot be used as a stopgap while the Enterprise contract is negotiated—the compliance features simply are not present. The review trail on major software review sites is also thin: one G2 review as of independent verification, so the evidence base rests more on named customers (Cursor, Cognition, HubSpot) and vendor-published benchmarks than on aggregated user reviews[reference:11].

Exa vs. Tavily and Brave Search API

This comparison uses official pricing pages, documentation, and independently published benchmark data available as of November 2026. It focuses on the structural differences that determine retrieval accuracy, throughput ceiling, and cost structure—not on declaring a universal winner. Exa is a semantic search API over its own index; Tavily is an AI-agent search API with a free credit model and a thinner public disclosure of its retrieval stack; Brave Search API is an independent index built for high-throughput, keyword-oriented retrieval. The three overlap at the search endpoint but diverge on what they index, how they price content, and how fast they let you go.

Comparison checked against official product information on 2026-11-04. Features and pricing may change.