Exa API 2026: The Search Engine Built for AI Agents, Not Humans

What Is Exa?

Exa is an AI-native search engine built by Exa Labs, a San Francisco applied-AI research lab that has been building search infrastructure for AI systems since 2021 (the org on GitHub is exa-labs, created 2021-01-24, with 108 public repos). Where Tavily, SerpAPI, and Serper are repackaged web search for LLM-era consumers, Exa maintains its own index, its own retrieval models, and its own crawlers — and structures the index around the kinds of queries that AI agents and RAG pipelines issue in production.

That orientation shows up in three concrete ways. First, the API is six endpoints, not one: /search for retrieval, /search with type=deep-lite|deep|deep-reasoning for multi-step research, /contents for clean page extraction from known URLs, /answer for an LLM answer with citations, /monitors for scheduled searches that surface new web content, and /agent for async multi-step research, list building, and enrichment. Second, the index is vertical-aware: a single category parameter selects from web, company, person, publication, and (legacy) github / pdf / tweet categories — the company and people results are enriched profiles spanning publications, top collaborators, research areas, and funding. Third, the August 2026 Dynamic Highlights release is structured around an agent-shaped problem: tokens are scarce, so Exa now picks excerpts across the result set rather than treating each page independently.

Verified Pricing: $7 per 1k Search, $0.012-$1.00 per Agent Run

Exa's pricing model is pay-as-you-go with no subscription, no minimum spend, and a generous free tier — and every number below was verified from the live pricing page on 2026-09-08.

Search — $7 per 1,000 requests (first 10 results), $1 per 1,000 extra results, $1 per 1,000 AI page summaries. The base request includes up to 10 results with token-efficient page contents; additional results above 10 bill at $1 / 1k, and Exa-generated AI summaries bill at $1 / 1k pages on top. That is the most common endpoint — most production agents route everything through /search and only escalate to /agent when the query is genuinely a multi-hop research task.

Deep Search — $12 to $15 per 1,000 requests by tier. type=deep-lite ~4 seconds at $12 / 1k for lightweight synthesis; type=deep 4-15 seconds at $12 / 1k for multi-step reasoning with structured outputs; type=deep-reasoning 12-40 seconds at $15 / 1k for hard research tasks. Pick the tier that matches your agent's latency budget — deep-lite for fan-out research, deep-reasoning only for the hard multi-hop questions.

Contents — $1 per 1,000 pages. For known URLs: clean LLM-ready web content extraction including JavaScript-rendered pages, PDFs, and complex layouts. Cheapest endpoint in the family and the right tool for ingestion into a RAG pipeline when you already know which pages you want.

Answer — $5 per 1,000 requests. A direct LLM answer to a question with citations — useful when the agent just needs a number or a fact and the URL of the source it came from.

Monitors — $15 per 1,000 requests. Scheduled searches that surface new events on the web. $1 / 1k results above 10 plus $1 / 1k AI summaries on top.

Agent — $0.012 to $1.00 per fixed-effort run, or usage-based. Pick a fixed effort on /agent for a predictable per-request price: minimal $0.012, low $0.025, medium $0.10, high $0.50, xhigh $1.00. Metered runs (the auto default and the beta max mode) bill actual usage: ACUs at $0.10, search tool calls at $0.005, email enrichment at $0.02, phone enrichment at $0.07. auto defaults to a $5 cap per run; beta max defaults to a $20 cap. Connect providers (Fiber $0.02/credit, Baselayer $0.15-$4.00/order, etc.) bill separately per call.

Free tier — $20 in credits at signup, $10 every month, no payment method required. That is roughly 2,800 searches to start plus ~1,400 searches of free credits per month. The free tier includes access to all six endpoints, not a stripped-down subset. Get a key at dashboard.exa.ai/api-keys.

Enterprise — custom via sales demo. Volume discounts, custom indexes, stricter security, Zero Data Retention, higher rate limits, in-VPC deployments, federated auth, log exports, audit logs, SLAs. A 14-day free trial is available. The honest caveat: Connect data sources bill unpredictably per task, and there is no mainland-China hosting region — direct access depends on local ISP egress.

Why Dynamic Highlights Is the Big August 2026 Release

Most search APIs return ten summaries, one per page, and your agent picks whichever seems most relevant. That is a fine heuristic for humans, who can read ten summaries in seconds — but for agents, every summary costs tokens, and the marginal value of the eighth or ninth summary is usually low because it is repeating what the first three already said.

Dynamic Highlights flips that: instead of treating each page independently, Exa selects excerpts across the complete result set. The agent's tokens go to the excerpts that add information not already covered, and not to excerpts that only repeat it. The mode is opt-in: send dynamic: true and the Exa-Beta: dynamic-highlights-2026-08-28 header.

Exa publishes the numbers from a representative benchmark, and they are large enough that any team doing RAG or agent retrieval should test this mode against their own workload. The published deltas: ~49% better token efficiency and +2.4% downstream quality for single-turn RAG (measured on Exa Auto across coding and general QA evaluations); ~30% fewer tokens across complete agent trajectories and +1% quality across BrowseComp, WideSearch, and internal company / people evaluations. Because the underlying harness is open source (exa-labs/benchmarks on GitHub), the numbers are reproducible — not just vendor-claimed.

For a team running RAG at scale, this is the single biggest feature to evaluate in 2026. The benchmark is real, the beta header is documented, and the rollout is one request parameter away.

Six Endpoints: Which One to Use When

The hardest decision is which endpoint to call. Exa gives you six, and they have meaningfully different trade-offs.

• Use /search by default. Real-time retrieval with token-efficient page contents. $7 / 1k for up to 10 results. Right for fan-out retrieval in RAG and as the first call in most agent loops.
• Use /search with type=deep when the agent's question is multi-hop. Deep Search performs multi-step reasoning with structured outputs and citations. Pick deep-lite for lightweight synthesis (~4s), deep for structured outputs (4-15s), deep-reasoning only for hard questions (12-40s).
• Use /contents when you already have the URLs. Cheapest endpoint ($1 / 1k pages) for ingesting known URLs into a RAG pipeline. Handles JS-rendered pages, PDFs, and complex layouts.
• Use /answer when the agent just needs a number. Direct LLM answer with citations at $5 / 1k — best for Q&A and fact lookup where the URL of the source matters more than the source's full text.
• Use /monitors for new-event pipelines. Scheduled searches that surface new content — $15 / 1k for the base request, plus extras. Right for news monitoring, GTM enrichment, and lead nurturing workflows.
• Use /agent for async multi-step research. Pick a fixed effort (minimal $0.012 through xhigh $1.00) for predictable per-request pricing, or metered mode ($5 default cap, $20 beta max cap) for usage-based billing.

The practical mental model: search retrieves, deep search reasons, contents extracts, answer summarises, monitors tracks, agent does everything asynchronously with structured outputs. Use the cheapest endpoint that solves the question.

MCP-Native: Hosted + Self-Hosted MCP Server

MCP support is the headline deployment story for any 2026 search API. Exa ships two MCP paths.

Hosted MCP server. The endpoint is https://mcp.exa.ai/mcp. It accepts POST with an MCP initialize handshake and serves a working MCP stream; HEAD returns 405 (which is the correct answer for a server expecting POST). The hosted path lets you add Exa to Claude Desktop, Cursor, Codex, VS Code (with the MCP-enabled extension), or any MCP-compatible client without running any infrastructure — the agent gets /search, /get_contents, /research, /company_research as tools, and the free-tier API key is enough for evaluation.

Self-hosted MCP server. The repo exa-labs/exa-mcp-server has 4,987 GitHub stars, 383 forks, MIT-licensed, last commit 2026-08-21 (publish container manifest as a single-arch Docker image). It exposes the same tools and is the right path if you want to keep API keys inside your own infrastructure, or if you want to extend the tool surface with internal Exa-powered research routines.

Both paths integrate cleanly with the rest of the AI tooling stack. Exa publishes integration pages for Vercel AI SDK, LangChain, CrewAI, LlamaIndex, Pydantic AI, n8n, and OpenAI Agents, plus a Stripe Projects integration that lets a coding agent sign up, provision an API key, and start retrieving without human intervention — useful when your agent needs to bootstrap itself.

Open Source, Open Evals, Open Index Decisions

The single most distinctive thing about Exa in 2026 is that the vendor's benchmark is reproducible. Most search-API vendors publish speed numbers and recall numbers with no way to verify them — Exa publishes the harness as exa-labs/benchmarks on GitHub, including the public retrieval benchmark added in July 2026 alongside the 350M-publication expansion.

The same orientation shows up in the SDKs. exa-py (the Python SDK) is at 233 stars, MIT-licensed, with the most recent release 2.20.0 on 2026-09-01 and the most recent commit on 2026-09-03 adding Dynamic Highlights (research preview) via the betas parameter. exa-js is the official JavaScript / TypeScript SDK. Both follow the same versioning cadence and the same API surface as the REST API.

The other thing worth noting is the index openness. Exa continuously indexes the live web (the same crawlers that power the Search API), and the index is queryable in real time — there is no "next-day" staleness the way some enterprise crawlers have. The 350M-publication vertical added in July 2026 is a useful counter-example to the "search engines are stale on niche verticals" critique.

Exa vs Tavily vs Firecrawl vs SerpAPI vs Google Search

Exa publishes a full comparison matrix against every major competitor: Tavily, Parallel, Perplexity, Firecrawl, Google Search, Brave, SerpAPI, Serper, and You.com. The crisp positions:

• vs Tavily. Tavily is the closest competitor and ships as a hosted SaaS for agent builders. Exa wins on the index side (Exa's own crawlers, including the 350M-publication vertical) and on the per-tier predictability of the Agent effort pricing ($0.012-$1.00 vs Tavily's per-1k pricing). Tavily has a more polished agent-builder UI; Exa has a cleaner SDK and reproducible benchmark.
• vs Firecrawl. Firecrawl is the right tool for crawling known sites and converting to markdown — pair Firecrawl with Exa, do not pick one over the other. Use Firecrawl for site-scale ingestion, use Exa for live search and structured retrieval.
• vs SerpAPI / Serper. SerpAPI and Serper are repackaged Google SERP scraping. They are good proxies for "what would Google return"; Exa is a different product — its index is structured around AI-agent queries, not human queries, and Dynamic Highlights is built around agent-shaped token budgets. If your agent is doing RAG, Exa wins. If your agent is doing SERP-rank-checking, SerpAPI wins.
• vs Perplexity. Perplexity is an end-user AI search product, not an API. Exa is the closest underlying layer that Perplexity-like-style products are built from — many agent builders pick Exa as the retrieval layer for their own Perplexity-like product.
• vs Google Search (Custom Search JSON API). Google's API is human-shaped SERP results and is the most restrictive of the comparison set (quota, billing friction, no agent-shape features). Exa is a better fit for any agent workload.

Who Should Use Exa?

Exa is the right fit when you tick at least two of these boxes:

• You are building an AI agent that needs the live web. Retrieval-augmented generation, company / people research, GTM enrichment, lead qualification, news monitoring, academic publication search. Exa's six endpoints and the August 2026 Dynamic Highlights release are built exactly for these workloads.
• You need MCP-native integration with AI IDEs. Coding agents (Cursor, Codex, Claude Code) and AI IDEs (Claude Desktop, VS Code with MCP) need a hosted or self-hosted MCP server. Exa ships both.
• You want token-efficient retrieval. The published 49% token-efficiency benchmark for Dynamic Highlights is the most concrete proof that Exa treats agent context budgets as a first-class concern. If your agent is cost-sensitive, this matters.
• You want reproducible vendor benchmarks. exa-labs/benchmarks is on GitHub. You can run the benchmark against your own workload before committing.
• You want pay-as-you-go with no subscription. No monthly minimum, no seat-based pricing, no per-seat fee — just per-request billing with $20 free credits at signup to evaluate.

Exa is the wrong fit if your workload is bulk SERP-rank checking (pick SerpAPI), if you only need static-site-to-markdown conversion (pick Firecrawl), or if you need a self-hosted solution that runs entirely inside your own infrastructure (pick an open-source alternative — Exa is hosted SaaS only).

Final Verdict

Exa in 2026 is the most interesting single-purpose retrieval API on the market for AI agents. The combination of a vendor-owned index with 350M publications, six purpose-built endpoints, pay-as-you-go pricing with a $20 free tier, the hosted MCP server, the self-hosted MCP server with 4,987 stars, the August 2026 Dynamic Highlights release that demonstrably saves 30-49% of agent tokens, and integrations across OpenAI Agents, Vercel AI SDK, LangChain, CrewAI, LlamaIndex, Pydantic AI, and n8n is hard to match. The published benchmarks are reproducible on GitHub. The free tier is real (2,800 searches to start, 1,400 per month ongoing).

If you are building an AI agent that needs the live web, evaluate Exa before Tavily, SerpAPI, or Firecrawl — the per-tier pricing predictability, the MCP-native path, and the Dynamic Highlights token-efficiency work are the differentiators that matter for production workloads. If you are already running Tavily or SerpAPI and hitting token-cost or freshness walls, Exa is the upgrade path. The migration story is straightforward: the exa-py and exa-js SDKs are first-class, the API is single-endpoint https://api.exa.ai with Bearer-token auth, and the $20 free credits at signup are enough to migrate any agent workload end-to-end.