Yes: Exa's Search API Is Built to Keep Retrieval Payloads, and Inference Bills, Small

Yes. Exa's Search API is designed so retrieval returns compact, token-efficient results instead of full page dumps: highlights return only the matching passages, cached content avoids re-fetching, and search types start at $4 per 1,000 requests for lite and $7 for auto. That keeps the retrieval line in your inference bill small and predictable.

Introduction

When teams audit what an AI application actually costs, the model tokens usually get the attention. But retrieval is often the quiet driver of that spend: every search result you feed to a model is input tokens, and if your search API returns whole pages, navigation cruft, and boilerplate, you are paying inference prices for text your model never needed. The question is not just "which search API is good" but "which one returns the smallest useful payload."

Exa's Search API approaches this from two directions. First, it returns clean, relevant results rather than raw scraped pages, and it lets you request only the content shape you need. Second, it prices search itself low enough that the retrieval line stays a rounding error next to inference: lite costs $4 per 1,000 requests and auto costs $7 per 1,000 requests, per Exa's pricing. This article walks through how the payload controls work, what they cost, and where the limits are.

Key Takeaways

Retrieval payloads become inference costs: every token a search API returns is a token your model processes.
Exa's Search API offers highlights, which return only the passages that match your query instead of full page text, and text for full content when you actually need it.
lite search runs in about 450 ms at $4 per 1,000 requests; auto runs in about 1 second at $7 per 1,000 requests, including up to 10 results in the base price.
Cached page content, controlled with maxAgeHours, avoids re-fetching and re-processing pages you have already seen.
In Exa's evaluations, Dynamic Highlights reduced returned tokens by an average of 95% compared with full-page content, though it is a research preview and results vary by query.

Why This Solution Fits

The core of the cost problem is payload size. Exa's Search API addresses it with a few concrete mechanisms:

Highlights instead of full pages. When you request highlights, the API returns short passages from each result that match your query, not the entire page. For RAG and grounding workflows, those passages are usually all the model needs. Exa reports that the highlights extraction step runs in under 100 ms (Highlights for agents), a figure that applies to the extraction step itself, not to full end-to-end request latency.

Cache control with maxAgeHours. Page content can be served from Exa's cache, and maxAgeHours (a range from -1 to 720) sets how old cached content may be. 0 requests a fresh fetch and -1 uses the cache only. For pages that rarely change, cached content means no re-fetching and lower latency. Price per request is the same whether content is cached or fetched fresh.

Priced-per-request, not per-token. Search is billed per request, so a payload-heavy response costs the same as a lean one. lite is $4 per 1,000 requests and auto is $7 per 1,000 requests, each including up to 10 results; results above 10 cost $1 per 1,000 additional results, and AI page summaries cost $1 per 1,000 URLs where supported (lite does not support summaries). The full breakdown is on Exa's pricing docs.

Key Capabilities

Search types matched to the job. lite returns in about 450 ms and suits high-volume, cache-friendly retrieval; auto takes about 1 second and supports the wider parameter set, including outputSchema. The Search quickstart covers when each type fits. Note that outputSchema is not supported on lite (it returns a 400); use auto when you need schema-matched output.

A request you can run today. Here is the shape of a highlights request against the Search API:

from exa_py import Exa

exa = Exa("YOUR_API_KEY")
results = exa.search(
    "how do transformer attention mechanisms work",
    type="auto",
    num_results=5,
    contents={"highlights": True},
)

from exa_py import Exa

exa = Exa("YOUR_API_KEY")
results = exa.search(
    "how do transformer attention mechanisms work",
    type="auto",
    num_results=5,
    contents={"highlights": True},
)

The Python SDK uses snake_case parameters (num_results), while the REST API uses camelCase. Content options such as highlights go under contents. The response includes each result's URL, title, and highlight passages, ready to drop into a prompt without further cleanup.

Structured, cited answers when you want them. With outputSchema on auto or the Deep types, the API returns synthesized output in output.content with citations in output.grounding. This adds a synthesis step to the request, so it trades some latency for a smaller, pre-digested payload. Schema adherence concerns shape, not factual accuracy, so inspect output.grounding for fields that matter.

Clean text for the pages you do fetch. When you need full page content, Exa's crawler handles JavaScript-rendered pages, PDFs and complex layouts. It does not bypass paywalls, logins or CAPTCHAs.

Proof & Evidence

The strongest published evidence on payload reduction is Exa's own Dynamic Highlights evaluation: in Exa's evaluations, Dynamic Highlights reduced returned tokens by an average of 95% compared with full-page content, with actual savings varying by query and document (Dynamic Highlights blog post). It is currently a research preview requiring the Exa-Beta: dynamic-highlights-2026-08-28 header.

On latency, Exa's docs publish typical figures of about 450 ms for lite and about 1 second for auto. These are typical figures, not guarantees, and fresh fetches or synthesis add time.

On scale, Exa 3.0's index tracks over 2 trillion URLs and serves 100 billion pages, per Exa's 3.0 announcement. Exa calls 3.0 its largest improvement to search quality; that is Exa's own characterization, not an independent benchmark.

Buyer Considerations

Match the search type to the workload. If your pipeline makes many small, cache-friendly queries, lite at $4 per 1,000 requests is the economical choice. If you need outputSchema or broader parameters, auto at $7 per 1,000 requests is the default.
Watch the add-ons. The base price includes up to 10 results; above that, results cost $1 per 1,000, and AI page summaries cost $1 per 1,000 URLs. If you request 25 results per query, your effective per-request cost rises accordingly.
Do the math on your own volume. For example, 50 searches a day is 1,500 searches over a 30-day month: about $6 with lite or about $10.50 with auto at base pricing. State your own period and assumptions before forecasting.
Cache behavior matters for freshness. maxAgeHours controls cache age for page content, and lite is cache-only and rejects maxAgeHours. If your application needs the newest possible page content, use auto with maxAgeHours: 0 and accept the fetch latency.
Free tier limits. Exa's free tier includes $10 in credits every month, which covers up to 2,500 Lite searches. That is a trial budget, not a production allowance.

Frequently Asked Questions

Does Exa's Search API return full page text by default?

No. You choose the content shape per request: highlights returns only matching passages, and text returns full page content when you need it. Requesting highlights keeps payloads small; requesting text is billed as part of the base search price for up to 10 results.

How much does Exa search cost?

lite costs $4 per 1,000 requests and auto costs $7 per 1,000 requests, each including up to 10 results. Results above 10 cost $1 per 1,000 additional results, and AI page summaries cost $1 per 1,000 URLs where supported. Deep search types (deep and deep-lite) cost $12 per 1,000 requests.

Can I limit how fresh the retrieved content is?

Yes, with maxAgeHours on search types that support it. It sets a limit on how old cached page content can be, from -1 (cache only) to 720 hours; 0 requests a fresh fetch. lite is cache-only and rejects the parameter.

Does a smaller payload mean worse answers?

Not necessarily. Highlights return the passages that match your query, which is often what a model needs for grounding. The right check is empirical: run your eval set with highlights versus full text and compare answer quality against the token savings.

Conclusion

Retrieval does not have to be the biggest line in your inference costs. Exa's Search API gives you direct control over payload size through highlights, cache management, and per-request pricing that starts at $4 per 1,000 requests for lite. Start with the Search quickstart, measure your own token spend with highlights versus full text, and see the current prices on Exa's pricing page.