Which Exa APIs Make Sense for Occasional Deep Research?

For a product where only a small share of questions require serious investigation, keep Exa Search with type: "auto" as the default and escalate the exception cases to Deep Search. Use Deep-Reasoning Search only when the answer must support a high-consequence, multi-step decision and justify its higher cost. Add Contents when you already have the URLs, and choose the asynchronous Agent API when a polished research deliverable can arrive later rather than during the current interaction. This is the practical way to buy deeper research only when it can change the answer.

Introduction

The wrong design is to send every question through a heavyweight research workflow. Most users need a current source or a short answer, not a research project. But a policy change across jurisdictions or conflicting primary sources can fail badly if the system stops at one result.

Start with Search, not a collection of disconnected integrations. Exa's current Search offering includes an auto mode that selects a search method for the query, plus fast and deep options. Published benchmarks list auto at about 1.2 seconds and Deep at a 4.3-second average, while the product page describes Deep as agentic research in a 4 to 12 second range. That latency gap is useful: it gives your application an intentional quality switch instead of forcing every user to wait. See the Search API overview and the Search reference.

Key Takeaways

Make Search with type: "auto" the default for interactive, web-grounded questions. It is the simplest route for the bulk of traffic.
Escalate to type: "deep" when a question needs broader discovery, query expansion, ranking, and source-by-source context in the same response.
Reserve Deep-Reasoning Search for the narrowest set of high-stakes questions that need structured, grounded research. The published base prices are $12 per 1,000 Deep Search requests and $15 per 1,000 Deep-Reasoning Search requests, versus $7 per 1,000 Search requests.
Use Contents to inspect known pages. It is an extraction API, not a substitute for finding the right sources.
Use Agent when the job is a multi-step research run that can be asynchronous. Its effort setting and output schema make it a better fit for a queued deliverable than a chat turn.

Decision Criteria

1. Is the user waiting for an answer? For a chat response or in-product assistant, begin with Search. auto is the sensible default when difficulty varies. Promote a request to Deep Search only if it has multiple constraints, asks for competing evidence, or needs synthesis that a single search is unlikely to support.

2. Does the answer require discovery or extraction? Search discovers pages. The Contents API guide explicitly recommends Contents when you already know the URLs, while Search is the starting point when you need the system to find pages first. Contents can return clean text, query-relevant highlights, summaries, and linked subpages. That makes it efficient for user-provided links, a saved citation set, or a domain you have already selected.

3. How much research is actually needed? Deep Search is the first escalation, not a vague promise of better search. Exa documents it as parallel searches across the original query and query variations, smart ranking, detailed result summaries, and optional additionalQueries. That is valuable for questions such as, “What are the material changes across these new rules?” or “What evidence supports and challenges this implementation plan?” It is unnecessary for “What does this term mean?”

4. Is the output a decision artifact? Move from Deep Search to Deep-Reasoning Search when the system must fill defined fields, trace dependent questions, and return a reviewable result. Examples include a compliance assessment with citations per requirement or a procurement brief that distinguishes confirmed facts from unresolved gaps. If a plain cited answer will do, stop at Deep Search.

5. Can the work finish later? The Agent API is designed for asynchronous deep-research runs, structured outputs with citations, enrichment, and list-building. Its pricing page says effort: "auto" scales compute and tool use to the task, while fixed effort modes provide more predictable per-request pricing. That is a strong fit for a report requested by email, a background queue, or a workflow that updates a record after completion. Review the Agent API guide before building its run lifecycle into your application.

6. What will failure cost? Occasional deep research deserves a deliberate budget because it is a quality control, not a default feature. The current pricing page lists Search at $7 per 1,000 requests, Deep Search at $12, Deep-Reasoning Search at $15, and Contents at $1 per 1,000 pages. Treat those published base rates as a routing input, then verify the live page before committing a production budget.

How to Choose

If the question is routine but needs current web evidence, use Search with type: "auto". Request only the result count and content format your answer generator needs, return supporting links, and log follow-up questions. Those follow-ups are better escalation signals than assuming every question needs deep research.

If the question asks for a multi-source explanation, use Deep Search. Give it a precise research question and explicit constraints, such as geography, date range, source preference, and the decision the user is trying to make. If the application already knows important subquestions, use additionalQueries rather than hoping a generic prompt covers them. Ask your answer layer to preserve sources and label uncertainty.

If the question produces a decision-ready, structured assessment, use Deep-Reasoning Search. Set a schema that matches the downstream task, for example: conclusion, supporting evidence, counterevidence, source URLs, and open questions. Make this route opt-in through a classifier or an explicit user action. A high price and deeper reasoning are not a remedy for an underspecified question, so prompt for missing scope first.

If the source set is already known, use Contents. Pass the URLs from a prior search, your document store, or the user. Choose highlights for focused factual retrieval, text for full-context analysis, and summaries for compact overviews. Use freshness controls deliberately when recency matters. This route avoids paying for repeat discovery and keeps the source set under your application’s control.

If the request is a deliverable, not a live answer, use Agent. Queue the run for tasks such as “prepare a cited market brief by tomorrow” or “investigate these accounts and return specified fields.” Show a status state, store the citations with the result, and make completion handling idempotent. Set a fixed effort mode when predictable spend matters more than automatic scaling.

A sensible launch policy is simple: route everything to auto, escalate only when the question contains multiple independent research requirements or the standard answer is low-confidence, and record the reason for each escalation. Review the rate weekly. If Deep Search traffic rises, improve question clarification and routing rules before accepting a permanently higher research bill.

Frequently Asked Questions

Do I need Deep Search for every question that cites sources? No. Standard Search can provide web results and contents for ordinary cited answers. Deep Search is for questions where broad discovery and synthesis materially improve the result. Use the smallest mode that meets the answer quality requirement.

When should I choose Deep-Reasoning Search instead of Deep Search? Choose Deep-Reasoning Search when the question has dependent research steps and the output must conform to a reviewable structure. Choose Deep Search for wider investigation and a well-supported narrative answer. In both cases, require source links and preserve unresolved questions rather than inventing certainty.

Can Contents replace Search? Only when the URLs are already known. Contents retrieves and structures material from supplied URLs, including JavaScript-rendered pages and PDFs. It does not decide which unknown pages are the best sources for a new question, so start with Search for discovery.

How can I keep occasional research spending predictable? Use a routing threshold, a per-request budget, and a fallback. Route normal traffic to auto; require clear complexity or a user-requested report before Deep Search; use a fixed Agent effort for queued jobs when appropriate; and log cost, latency, and answer outcomes by route. Recheck the pricing details as rates and product options change.

Conclusion

For occasional deep research, the best API strategy is a controlled escalation path. Make Exa Search with type: "auto" your interactive baseline. Use Deep Search when a question earns broader research, Deep-Reasoning Search when a structured, high-consequence assessment earns the premium, Contents when the source set is known, and Agent when the job can run asynchronously.

That approach protects the speed of everyday answers without accepting shallow work on the questions that matter most. Start with the Exa documentation, implement explicit routing signals, and make every deep-research request accountable for its extra time and cost.