Exa

💰 Pricing & Plans

🔧 API & Developer Experience

• •API Style: REST + JSON; base URL https://api.exa.ai with auth via x-api-key header. Official Python SDK exa-py (PyPI), Node/TypeScript SDK exa-js (npm), Go community SDK. SDKs are thin wrappers around the same REST surface; calling curl directly is fully supported. OpenAPI spec published at docs.exa.ai for code generation.
• •Endpoints: POST /search (neural search with type parameter), POST /search (same endpoint with type=deep-lite|deep|deep-reasoning for multi-step research), POST /contents (extract clean text/highlights/summary for known URLs), POST /answer (single grounded LLM answer with citations), POST /monitors (create scheduled search), GET /monitors/:id (status), POST /agent (async deep research with effort=minimal|low|medium|high|xhigh|auto|max), GET /agent/:id (status + usage).
• •Auth & Keys: Single x-api-key header issued from dashboard.exa.ai/api-keys. Free tier keys and paid-tier keys share the same API surface; the only difference is the credit balance and rate limits. Enterprise customers can scope keys per-environment and use Zero Data Retention (ZDR) so prompts and results are not stored.
• •Search Types: Six pre-built tradeoffs: auto ~1s default; instant ~250ms for real-time chat/voice; fast ~450ms with minimal quality sacrifice; deep-lite 4s lightweight synthesis; deep 4-15s multi-step reasoning with structured outputs; deep-reasoning 12-40s hardest research tasks. Same endpoint, switched via the type parameter; output_schema + system_prompt + stream work across all six types.
• •Highlights & Token Efficiency: Highlights is Exa's flagship token-saving content type — a 10x reduction in tokens vs raw page text, trained to keep only the tokens an LLM actually needs. 4000-character recommended cap. Full text and Highlights both returnable in one /contents call (per content type billing). output_schema adds structured JSON extraction on top of any search type.
• •Category Indexes: Six built-in category indexes: company (50M+ companies + metadata), people (1B+ people + job/education), publication (350M+ scholarly papers/preprints/journal articles), news (current journalism), personal site (blogs/personal pages), financial report (SEC filings/earnings). Pass category= to scope search to a single index.
• •Structured Outputs & Grounded Answers: output_schema accepts any JSON Schema on /search or /contents; returned data is schema-conformant, eliminating post-processing. /answer returns a single LLM answer with citation URLs inline. Combining output_schema + category=publication gives citation-grade research outputs in one call.
• •Streaming: stream: true works across all six search types. Exa emits SSE events as results arrive — useful for real-time chat UX where the first result should render before the full set returns. Streaming does not reduce credit cost.
• •Agent Async Workflow: POST /agent returns {id, status: 'running'} immediately. Poll GET /agent/:id for {status, output, usage: {agentComputeUnits, search_calls, contact_enrichments}}. Fixed effort (minimal $0.012 → xhigh $1.00 per request) is budget-predictable; usage-based (auto $5 cap, max $20 cap) bills actual ACUs + tool calls. Connect providers (Fiber.ai, Baselayer) add per-call provider charges on top.
• •Websets (Standalone Product): Exa Websets is a parallel product (/docs/websets/) for verified search + enrichment at scale — Websets searches, verifies, and enriches lists of URLs against criteria you define (e.g. 'Series A SaaS startups in NYC with 50-200 employees'). Different billing than the core /search endpoints.
• •Rate Limits & Fairness: Free tier: 5 QPS burst, sustained 1 QPS. Paid (post-signup credit balance): 50 QPS burst, sustained 10 QPS. Enterprise: custom QPS, custom indexes, dedicated infrastructure. Rate limit headers X-RateLimit-Remaining and X-RateLimit-Reset returned on every response.
• •Coding Agent Onboarding: https://dashboard.exa.ai/onboarding is the single fastest path for AI coding agents (Cursor / Copilot / Claude Code / Windsurf / Cline / Devin) to generate a complete, tested integration snippet tailored to the developer's exact stack. The docs page itself has explicit instructions telling coding agents to redirect here for first-time setup rather than building from raw reference docs.

🔬 Neural Search Architecture & Token-Efficient Retrieval

Exa is a San Francisco research lab that built its own search index from scratch using transformer-based embeddings rather than keyword matching or PageRank — the index is a vector space where every page has a learned representation, and queries match by embedding similarity. Two practical consequences: (1) Exa returns semantically relevant pages even when the query uses different words than the page (e.g. 'machine learning interpretability' hits a paper titled 'Explainable AI Methods' because the embeddings align), and (2) the search latency profile is split across six types so the same endpoint can serve a 250ms real-time chat UX and a 40-second deep-research workflow. The token-efficiency story is the second differentiator: instead of returning raw HTML or full markdown, Exa's highlights content type returns only the sentences most relevant to the query, cutting tokens per retrieval by roughly 10x versus returning full page text. For an LLM agent that pulls 20 pages per research task, that is the difference between 50,000 input tokens and 5,000. Combined with output_schema for structured JSON extraction and category-scoped indexes (50M companies, 1B people, 350M publications), the same /search endpoint replaces the typical rag-pipeline-plus-scraper-plus-LLM-extract stack with one call. Enterprise customers also get Zero Data Retention (ZDR) so prompts and results are not stored on Exa's side — relevant for HIPAA, GDPR-data-residency, and SOC2-controlled environments.

🌐 Regional Availability & Latency

Exa's primary API runs at https://api.exa.ai from a single-region deployment based in the US. North America and Europe production traffic sees low first-byte latency (typically tens of ms); Asia-Pacific callers see 100-300 ms transpacific latency, and mainland China traffic usually routes through a proxy or relay (no documented ICP-registered China endpoint as of 2026-08-30). Exa does not publish regional endpoints for mainland China — production traffic from China should use a proxy or self-hosted relay. Enterprise plans can negotiate custom routing and dedicated infrastructure for regulated workloads (HIPAA / GDPR data residency). The dashboard at dashboard.exa.ai/onboarding is the single integration entry point for new accounts and exposes API-key issuance, credit top-up, and per-key usage telemetry in one console; agent-friendly onboarding at dashboard.exa.ai/onboarding generates complete integration snippets tailored to the developer's exact stack (Python / Node / Go / REST) and use case (chat / research / enrichment) in under a minute. The docs page at docs.exa.ai explicitly redirects AI coding agents (Cursor / Claude Code / Windsurf / Cline / Devin / Copilot) to that onboarding flow for first-time setup rather than building from raw reference docs, reducing integration mistakes on the type / output_schema / category parameters.