Exa
Listed at https://exa.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | $7 / 1k Search requests (first 10 results; +$1 / 1k extra results; +$1 / 1k AI page summaries) | per million tokens |
| Output | Deep Search: $12-$15 / 1k requests • Agent: $0.012-$1.00 per fixed-effort run • Contents: $1 / 1k pages • Answer: $5 / 1k • Monitors: $15 / 1k | per million tokens |
🤖 Supported Models (7)
✨ Pros
- ✓Exa Labs (San Francisco applied-AI research lab) self-built search engine purpose-built for AI agents — real-time web index plus structured data on companies, people, financial markets, code, and documents
- ✓Six core APIs: Search ($7/1k), Deep Search ($12-$15/1k), Contents ($1/1k), Answer ($5/1k), Monitors ($15/1k), Agent ($0.012-$1.00 per fixed-effort run)
- ✓Agent API five-tier predictable pricing: minimal $0.012 / low $0.025 / medium $0.10 / high $0.50 / xhigh $1.00 per request
- ✓$20 signup credits + $10/month free credits, no payment method required, free tier includes access to all endpoints (~2,800 searches to start)
- ✓Official MCP server (exa-labs/exa-mcp-server, 4,987 ⭐) — works with Claude, Cursor, Codex, VS Code, and other MCP-compatible clients
- ✓Dynamic Highlights launched 2026-08-28 (research preview): ~49% better token efficiency for single-turn RAG, ~30% fewer tokens across agent trajectories with +1% quality
- ✓Comparison pages published vs Tavily / Parallel / Perplexity / Firecrawl / Google Search / Brave / SerpAPI / Serper / You.com
- ✓OpenAI-compatible friendly: integrates with OpenAI Agents, Vercel AI SDK, LangChain, CrewAI, LlamaIndex, Pydantic AI, n8n
⚠️ Cons
- ×Search is not a Google-style ranking — it uses Exa's own embeddings and retrieval models, and long-tail query quality diverges from traditional search
- ×No public batch / async uniform discount tier — Agent metered defaults cap at $5/run, beta max caps at $20/run
- ×Connect data sources (Fiber $0.02/credit, Baselayer $0.15-$4.00/order) bill separately per call; multi-source enrichment bills are unpredictable per task
- ×No mainland-China hosting region — direct access to the official SaaS depends on ISP and compliant egress, teams must self-evaluate latency and reachability
- ×Agent / Deep Search latency runs 4-40 seconds (deep-reasoning is the slowest tier) — not suitable for sub-second UI-real-time scenarios
🎯 Best For
Research teams that need real-time web retrieval and structured data for AI agents and RAG pipelines; workflow builders doing GTM, lead enrichment, company-and-people research, academic publication search, and new-event monitoring; developers who need native MCP integration with Claude, Cursor, and other AI IDEs; and indie devs and small teams who want a quick path to a production-grade search API with OpenAI-compatible interfaces and SDKs across Vercel AI SDK, LangChain, CrewAI, and LlamaIndex.
💰 Pricing & Plans
| Tier | Cost | Best For | Key Features |
|---|---|---|---|
| Free (no payment method) | $20 signup + $10/month credits | Evaluation, indie projects, hobby agents | Access to all 6 endpoints (Search, Deep Search, Contents, Answer, Monitors, Agent); ~2,800 searches/mo to start; dashboard.exa.ai/api-keys |
| Pay-as-you-go | Search $7/1k • Deep $12-$15/1k • Contents $1/1k pages • Answer $5/1k • Monitors $15/1k • Agent $0.012-$1.00/run | Production agents and growth-stage startups | All endpoints + extra-results billing ($1/1k results above 10) + AI summaries ($1/1k pages); no subscription, no minimum spend |
| Enterprise (sales demo) | Custom — volume discounts | High-volume workloads, custom indexes, Zero Data Retention | Higher rate limits, custom security controls, SLAs, in-VPC deployments, federated auth, log exports, audit logs; 14-day free trial |
🔧 API & Developer Experience
- •Single REST Endpoint with Six Endpoints: All requests go to https://api.exa.ai with a path per endpoint: /search, /search (with type=deep), /contents, /monitors, /answer, /agent. Bearer-token auth via EXA_API_KEY. Python SDK (exa-py, 233 ⭐ on GitHub, last release 2.20.0 on 2026-09-01) and JS/TS SDK (exa-js) are official.
- •OpenAI-Compatible for AI Tooling: Exa publishes integrations for OpenAI Agents, Vercel AI SDK, LangChain, CrewAI, LlamaIndex, Pydantic AI, and n8n — most let you drop Exa into an existing agent workflow by registering it as a tool. The official Stripe Projects integration lets coding agents sign up, provision an API key, and start retrieving without human help.
- •MCP-Native: Hosted + Self-Hosted MCP Server: Official MCP server at https://mcp.exa.ai/mcp (hosted, 405 only because it expects POST). Self-host from exa-labs/exa-mcp-server (4,987 ⭐, MIT-licensed). Works with Claude, Cursor, Codex, VS Code, and any MCP-compatible client — register Exa as an MCP tool and the agent calls /search, /get_contents, /research, /company_research automatically.
- •Structured Outputs + Token-Efficient Contents: Search and Deep Search both return clean LLM-ready contents with character-level highlights; set category='company'/'person'/'publication'/'github' (legacy categories pdf/tweet/github deprecated in Aug 2026) for vertical retrieval. AI page summaries are $1/1k pages on top of base request — opt in only when you need them.
- •Deep Search Latency Tier Table: Type=deep-lite ~4s / $12/1k, type=deep 4-15s / $12/1k, type=deep-reasoning 12-40s / $15/1k. Pick the tier that matches your agent's budget — deep-lite for fan-out research, deep-reasoning only for hard multi-hop questions.
- •Agent API with Predictable Effort Pricing: Set effort=minimal|low|medium|high|xhigh on /agent runs ($0.012-$1.00/request) for predictable per-request price. Metered mode (auto) bills ACUs ($0.10), search tool calls ($0.005/search), email enrichment ($0.02), phone enrichment ($0.07); default $5/run cap, beta max $20/run cap.
- •Dynamic Highlights (Aug 2026 Research Preview): Set dynamic=true with header Exa-Beta: dynamic-highlights-2026-08-28 to get highlights selected across the whole result set instead of per-page. Exa's benchmarks: ~49% better token efficiency and +2.4% downstream quality for single-turn RAG; ~30% fewer tokens across agent trajectories with +1% quality.
- •Publications Category + Public Retrieval Benchmark: July 2026 added a 350M-publication search category with richer org/people results. Public retrieval benchmark released on GitHub (exa-labs/benchmarks) — vendor evaluations are reproducible, not just vendor-claimed.
⚡ Why Exa: Purpose-Built Search for AI Agents
Most search APIs are repurposed web crawlers with an LLM bolted on afterward — they return ten blue links, hand them to your agent, and hope the LLM figures it out. Exa is built the other way around: the company describes itself as an applied AI research lab that builds its own index and search models, and the index is structured around the queries that AI agents actually make. The August 2026 Dynamic Highlights release is the clearest expression of that orientation — instead of returning ten per-page summaries, Exa now selects excerpts across the result set so the agent's context budget goes to useful sources and away from sources that only repeat information already returned. The benchmark Exa publishes (49% better token efficiency, 30% fewer tokens in agent trajectories, +1-2.4% downstream quality) is reproducible on GitHub. Combined with the six-purpose API surface (Search, Deep Search, Contents, Answer, Monitors, Agent), the hosted MCP server, and the integrations with every major AI tooling SDK, Exa is the closest thing the AI-agent ecosystem has to a default web retrieval layer — and it is priced per request, not per seat, with $20 free credits at signup.
🌐 Regional Availability & Latency
Exa is hosted SaaS, not self-hosted. The official api.exa.ai endpoint serves from Exa's own infrastructure (Next.js + Vercel per the public marketing site headers, with the underlying indexers running on Exa's own crawlers). Verified live on 2026-09-08: exa.ai returns HTTP/2 200 via Vercel edge; api.exa.ai accepts Bearer-token POSTs against /search, /contents, /monitors, /answer, /agent; mcp.exa.ai/mcp returns 405 on HEAD (which is the correct answer for an MCP server expecting POST) and serves a working MCP stream on POST. Mainland China teams should expect direct access to depend on local egress — there is no official mainland-China region, but since Exa is API-only with no GUI requirement, teams typically route via a compliant egress or call from a non-mainland worker. Deep Search tier latency runs 4-40 seconds (deep-reasoning slowest), which is dominated by Exa's own multi-step retrieval and LLM reasoning, not network round-trips. Agent API defaults cap a single run at $5 (auto) or $20 (beta max), so a runaway multi-source enrichment job is bounded.