Exa

Exa is a search API, not a search engine you visit — the product is /search, /contents, /answer, /monitors, and /agent endpoints that a developer calls from code. It runs its own crawl and index (Exa says it tracks over 500 billion URLs) with embedding-based retrieval rather than wrapping Google or Bing SERPs, and returns results pre-trimmed to the passages an LLM actually needs, which is the pitch for token cost. It ships official Python (exa-py) and TypeScript SDKs, plus a hosted MCP server (mcp.exa.ai) that plugs into Claude, Cursor, ChatGPT/Codex, and a dozen other agent clients without running any local server. Source: https://exa.ai/pricing, https://github.com/exa-labs/exa-py, https://exa.ai/docs/reference/exa-mcp

There's no monthly plan: you load credits and pay per call. Base search is $7 per 1,000 requests (covers up to 10 results, with page text and highlights included since a March 2026 pricing simplification), $1 per 1,000 additional results beyond 10, and $1 per 1,000 AI-generated page summaries. /answer (a direct LLM answer with citations) is $5 per 1,000 requests. /contents (fetching known URLs) is $1 per 1,000 pages per content type. Deep/multi-step search runs $12–15 per 1,000 requests depending on mode. The separate Agent API bills per fixed-effort run ($0.012 to $1.00) or by metered usage (compute units, tool calls, contact-enrichment lookups) with a default $5 cap. New accounts get $20 in free credits (~2,800 searches) plus $10/month ongoing on the free tier. Enterprise tiers (custom indexes, higher rate limits, SLAs, zero data retention) are quote-only. Source: https://exa.ai/pricing (verified 2026-08-27)

Independent analysis (Sacra, January 2026) puts Exa's per-1k-search rate at $5–7 against SerpAPI-style Google/Bing SERP wrappers charging roughly $25 per 1,000 searches — a meaningfully different pricing model because Exa isn't reselling someone else's index. The trade-off: Exa's own index and ranking, unlike a SERP wrapper, means result quality is entirely Exa's engineering, not Google's. Source: https://sacra.com/research/exa-at-10m-growing-11x-yoy/

Exa publishes head-to-head accuracy tables against Perplexity, Brave, Tavily, and Parallel on FRAMES, Tip-of-Tongue, and Seal0 — established retrieval benchmarks (FRAMES is a published Google DeepMind benchmark, arXiv:2409.12941) — claiming 54.4% vs Perplexity's 44.5% and Brave's 21.6% on FRAMES, and 178ms p50 latency for its fastest tier against 260–3,051ms for competitors. These are real, named benchmarks, but the comparison numbers are Exa's own test runs, not scored by an independent leaderboard — we could not find a neutral third party reproducing these exact figures. Treat the benchmark names as legitimate and the specific scores as a vendor claim until corroborated. Source: https://exa.ai/ (homepage benchmark tables, verified 2026-08-27), https://arxiv.org/abs/2409.12941

Exa says it's used by over 500,000 developers and 5,000+ companies (its own figures, from the May 2026 Series C announcement). Named, verifiable customers with published case studies include Cursor, Cognition (Devin), HubSpot, Firefox, monday.com, OpenRouter, and CodeRabbit — these are on Exa's customer-story pages with attributed named quotes (e.g., HubSpot's AI Lead, Cognition's co-founder), which is a stronger form of evidence than a logo wall but is still vendor-published. Independent research firm Sacra estimated Exa's annualized revenue at $10M in September 2025, up from roughly $880K a year earlier — implying a roughly 70x revenue multiple at its Series B valuation. Sources: https://exa.ai/blog/announcing-series-c, https://exa.ai/customers, https://sacra.com/research/exa-at-10m-growing-11x-yoy/

Exa (Exa Labs) has raised in three public rounds: a $17M round reported by TechCrunch in July 2024 (participants included Lightspeed and Y Combinator), an $85M Series B led by Benchmark in September 2025 at a $700M valuation, and a $250M Series C led by a16z in May 2026 at a $2.2B valuation — all confirmed via Exa's own funding announcements and corroborated for the earlier round by TechCrunch. Total disclosed funding is roughly $352M. The company was founded in 2021 (per its own account) and has been operating and shipping monthly product updates continuously through August 2026. Sources: https://exa.ai/blog/announcing-series-c, https://exa.ai/blog/announcing-series-b, https://techcrunch.com/2024/07/16/exa-raises-17m-lightspeed-nvidia-ycombinator-google-ai-models/

Exa publishes a SOC 2 Type II attestation, a public vulnerability disclosure policy with a paid bug bounty ($25–$5,000 by severity) at exa.ai/security, and a Vanta-hosted trust center listing GDPR, CCPA, and HIPAA (BAA available) support. Zero data retention is available but gated behind an Enterprise plan and a sales conversation rather than a self-serve toggle — the pricing page and enterprise FAQ both confirm this is quote-only. Source: https://exa.ai/security, https://exa.ai/enterprise, https://trust.exa.ai

The Exa product itself (crawler, index, ranking models) is closed and only accessible via the paid API. What's open source are client libraries: exa-py (229 GitHub stars, MIT license, actively maintained with releases through the SDK's changelog) and exa-mcp-server (4,926 stars, MIT license, the more actively developed and popular of the two). Neither is the product — they're thin wrappers around the hosted API and offer no path to running Exa's search independent of Exa's servers. Sources: https://github.com/exa-labs/exa-py, https://github.com/exa-labs/exa-mcp-server (verified 2026-08-27)