Exa is a search index and retrieval API built for agents. The company describes itself as an applied AI research lab. It keeps its own index of the public web plus structured collections such as companies, people, filings, papers, and jobs. The homepage at exa.ai is a script-rendered shell. The page itself tells an agent to start at exa.ai/llms.txt. This note follows that file and the pricing page it links.
Checked 3 November 2026. No account was created and no API call was made. The homepage claim of 500,000 developers was not counted.
How a call is made
The pricing page says the API is served at https://api.exa.ai. Send the key as Authorization: Bearer $EXA_API_KEY or as x-api-key. Keys are created at dashboard.exa.ai/api-keys. llms.txt says /search and /contents also accept x402 pay-per-request payments. The OpenAPI files at exa-spec.yaml and team-management-spec.yaml are the schema source. The guides are usage advice, not the schema.
Official clients are exa-py (pip install exa-py) and exa-js (npm install exa-js). Both read EXA_API_KEY from the environment. A hosted MCP server at https://mcp.exa.ai/mcp can be tried without a key, at a limited rate. The skill install line on the pricing page is npx skills add exa-labs/agent-skills.
Products and prices
The pricing page says there is no subscription and no minimum. You load credits and pay per request. A new account gets 10 on the first of each month. No payment method is required. Finishing onboarding adds a one-time 10 covers up to 2,500 Lite searches, which matches the Lite rate below. The free tier can call every endpoint.
Each base price includes up to 10 results. Every result past 10 adds 1 per 1,000 pages.
| Call | Base price | Note on the pricing page |
|---|---|---|
/search type lite |
$4 / 1,000 requests | No separate summary meter |
/search type instant |
$4 / 1,000 | Legacy type, same price |
/search type auto or fast |
$7 / 1,000 | fast is legacy. Summaries are $1 / 1,000 pages |
/search type deep-lite |
$12 / 1,000 | About 4 seconds |
/search type deep |
$12 / 1,000 | 4 to 15 seconds |
/search type deep-reasoning |
$15 / 1,000 | Legacy. 12 to 40 seconds |
/contents |
$1 / 1,000 pages, per content type | text, highlights, and summary are billed apart |
/answer |
$5 / 1,000 requests | A model answer with citations |
/monitors |
$2 / 1,000 requests | A scheduled search for new pages |
The homepage labels latency as Lite 450 ms, Auto 1 s, and Deep about 10 s. Those are the homepage's labels. The ranges in the table are the ones next to the price.
The Agent API prices a run by an effort value. minimal is 0.025, medium is 0.50, and xhigh is 5 per run. ultra uses the same meters with a default cap of 0.10 per Agent Compute Unit, 0.02 per email found, and $0.07 per phone number found. Exa Connect adds the upstream provider's own price on top. Zero Data Retention, a custom index, more than 25 results on the self-serve plan, and an SLA are enterprise terms, not the default plan.
What else the index file lists
Exa Library is a set of structured collections queried through the same search call. The file names places, people, companies, job postings, earnings calls, SEC filings, stock pages, market quotes, products, events, podcasts, Wikipedia, research papers, patents, court opinions, indicators, countries, sports, and weather.
The file also lists comparison pages, including Exa vs Brave. Those pages are Exa's. None of them was opened, so none of their scores is repeated here. Published evals and an open benchmark repo are linked from llms.txt. The runs were not repeated.
The Cloudflare Web Search API note already records Exa as the exa provider at 7 figure matches the auto row above. The enterprise-only ZDR line matches that No.
See also: Brave AI · SearXNG · Cloudflare Web Search API