Exa

Web-grounded retrieval, then synthesis. Every request is charged — and asking for more (results, content types, a deeper search) charges more.

Use Exa for web-grounded retrieval, then synthesis.

Managed credits — every search, contents, answer and find-similar request is charged, and the request itself sets the price: 0.2 credits for a search of up to 10 results, 0.1 for an answer or a contents fetch, more when you ask for more. See Cost below.

type vs category

These are independent params:

type controls the search strategy: auto / fast / neural / deep.
category filters result type: company / people / news / financial report.

"type":"news" is a 422 error — for news results use "category":"news".

Cost

Exa charges for the request it accepts, not for the rows it returns — so a search that matches nothing is still a charged search, and the card price is the price of the base request:

Tool	Base cost	What raises it
`exa_search`	0.2 / request (up to 10 results)	each result beyond the first 10 · a per-result `summary` · `type` `deep` / `deep-reasoning`
`find_similar`	0.2 / request	each result beyond the first 10
`exa_answer`	0.1 / request	—
`exa_contents`	0.1 / request	each URL · each content type (`text`, `highlights`, `summary`)
`exa_agent`	charged on completion	the run's own effort and work — see below
`exa_agent_get`, `exa_agent_runs`	free	—

Keep numResults tight and ask only for the content types you will actually read.
One call, not two: a search with contents costs less than a search followed by a separate exa_contents on the same URLs.
Agent runs are charged when they complete, for what the completed run came to. Starting one is free, polling it is free however often you poll, and a run that ends failed or cancelled is not charged. A fixed effort caps what a run can reach; the default auto has no cap, so bound the work with an explicit effort and maxItems on the output schema when the spend needs to be predictable.
--dry-run any call to see its resolved base price before a bulk run.

Guidance

Treat exa_answer as the summarization layer, not the first retrieval step, when precision matters. Run exa_search / exa_contents first and synthesize after inspecting citations.
Use focused queries and small numResults during pilots; widen only if coverage is low.
For auditable outputs, inspect citations before synthesizing.

hyreflow tools execute exa exa_search \
  --payload '{"query":"series b devtools companies united states","numResults":5,"type":"fast"}'

hyreflow tools execute exa exa_contents \
  --payload '{"urls":["https://example.com"],"text":true}'

hyreflow tools execute exa exa_answer \
  --payload '{"query":"Summarize the top GTM signals from these results","text":true}'

In the JD → Shortlist play, exa.answer("direct competitors of {company}…") discovers competitor domains. Don't use find_similar on a homepage — it returns directories, not competitors.

type vs category

Cost

Guidance

On this page