Exa API 定价（2026）：费用与科研论文替代方案对比

Choose Exa for configurable web retrieval. Choose SciSpace for a product whose users need papers found, read, compared, and synthesized. Exa also offers cited answers and research agents, so this is not a links-versus-answers comparison.

For literature assistants and evidence-review tools, the SciSpace API is the closer fit: one task can cover discovery, reading, extraction and a cited report delivered as files. The SciSpace API is in early access and isn't open yet.

This guide covers exa api pricing, five shared evaluation questions, request examples, and migration choices. Examples illustrate interfaces, not a completed head-to-head benchmark.

What does each API return to your product?

Exa Search returns ranked results and page content; Answer returns synthesized text with sources; Agent orchestrates research and structured output. Choose the endpoint that matches your deliverable.

The SciSpace API has two parts. The Search API takes a research question and returns papers with their metadata; the Task API takes a research task, picks its own tools, and returns links to the files it created plus any supporting text answer.

Features at a glance

Exa entries follow its Search, Answer, and Agent references. SciSpace entries reflect the SciSpace API in early access, October 2026.

Can Exa search academic papers?

Yes. Exa Search supports category: "publication" for scholarly retrieval, and its research documentation describes publication-screening and extraction workflows. A literature evaluation should therefore include Exa Agent, not only Search.

SciSpace's distinction is the integrated literature workflow, not exclusive access to academic research.

What is Exa API pricing per 1,000 calls?

Published Exa rates checked October 21, 2026:

Exa's rate card also lists legacy search modes. Exa Agent defaults to metered auto effort, charged at $0.10 per Agent Compute Unit plus $0.005 per search tool call. The default cap is $5 per run, not a flat charge. Contact enrichment and connected data providers can add charges. These metered rates also apply to ultra, whose default cap is $20 per run. Fixed-effort prices apply only when you explicitly select a fixed effort.

For search, results above ten add $1 per 1,000 additional results. Thus 1,000 Auto searches requesting 20 results total $17 before summaries or other extras: $7 base plus $10 for additional results.

Exa's free allowance and rate limits

Exa provides $10 in monthly credits, plus a one-time $10 onboarding bonus under its eligibility rules. Monthly credits do not accumulate.

Default Exa limits include 10 queries per second for Search and Answer, 5 for Deep Search, and 5 Agent run starts per second with 50 active runs. Limits apply across a team's keys and can vary by plan.

For an exa ai search api pricing estimate, specify endpoint, result count, and content options. Budgeting a retrieval request is different from defining an acceptable research deliverable.

How should you compare the same five research questions?

Use the same question text and instructions for both providers. Fix the cutoff at December 31, 2025 so changing publication dates do not obscure the comparison.

Execution disclosure: No authenticated calls were run for this article. These are proposed tests, not observed answers or performance claims.

Append these identical instructions to every question: Use only primary papers published on or before December 31, 2025. Cite major claims, distinguish inference from reported findings, and write ‘not reported’ for missing evidence.

Audit source relevance, claim support, and missing fields. Do not treat citation counts as accuracy scores or compare an Exa Search response with a completed review as if they were equivalent tasks.

What do equivalent requests look like?

Use Exa Answer for the initial cited-answer comparison. For a second evaluation, compare Exa Agent with a SciSpace Task API deep literature review, keeping those results separate. Exa Agent can orchestrate searches and reading itself.

Set prompt to one table question plus the shared instructions. Repeat for all five when access permits, preserving the exact text and raw responses.

The Python example requires the requests package. The SciSpace side is described in plain language, since the API is in early access and isn't open yet.

Exa: request a cited answer

This shortened Python example preserves the documented /answer request. text: True requests source text for inspection; it does not prove that every generated claim is supported.

SciSpace: what you send

Send the same prompt to the SciSpace Task API as a plain-language task. It streams its plan and progress while it runs, then returns links to the files it created, such as a Markdown or Word report, plus any supporting text answer.

One task covers the whole workflow, and follow-up questions continue in the same thread. See the deep research API guide for related workflow context.

How do the responses appear side by side?

The following is a response-contract comparison, not fabricated research output. Exa Answer uses answer and citations; the SciSpace Task API returns links to the files it created plus any supporting text answer.

Exa Answer supports outputSchema for structured output. The SciSpace extraction table has AI-filled columns from full text, such as datasets, methods, sample sizes or outcomes, which you can filter, sort and combine.

For both interfaces, keep evidence inspectable. Require paper identifiers, check extracted values against their sources, and preserve missing values rather than guessing.

For extraction design, retain the PDF data extraction guide alongside your acceptance tests.

When should you choose Exa instead?

Choose Exa when you need control over retrieved URLs, domain filters, and page content feeding an existing synthesis pipeline. Its Search interface suits that architecture.

Choose Exa Agent when your main requirement is general research with schema-defined output. It can coordinate searches and reading, so it would be inaccurate to say every Exa integration requires custom agent orchestration.

For literature products, the decision is whether you want to own the remaining paper-specific workflow. An existing extraction pipeline may justify retaining Exa; a new evidence-review product has stronger reasons to choose SciSpace.

For architectural context, see RAG APIs for research.

How do you migrate parameters from Exa to SciSpace?

This is not a base-URL replacement. Move the task into plain language and keep validation in your application.

Exa fields are endpoint-specific. Do not assume Search filters also apply to Answer.

A requested paper count is not a hard result limit, and a requested table layout is not schema enforcement. Preserve your application's validation when migrating.

The SciSpace agent does not pause mid-run for user input, so put screening criteria and research questions in the task itself. Show the streamed plan and progress in your interface, store the returned file links, and send refinements as follow-ups in the same thread.

The SciSpace API for literature research products

Exa searches the open web plus a 350M-publication index, and its Agent reports live progress, but it returns text and JSON rather than files. Its publication results carry DOI, authors, year, abstract and citation count, with no journal or study type (checked against Exa's API docs on 21 Oct 2026).

The SciSpace API puts the full SciSpace Agent behind one call. It searches, reads, analyses and writes, and delivers the result as ready-to-use files.

What this adds on top of Exa

• One integration instead of a pipeline: describe the task in plain language and the agent picks the tools (search, screening, extraction, analysis, writing). You do not stitch together a search API, a PDF parser and an LLM.
• Finished files, many from one run: Markdown, Word, PowerPoint, or a LaTeX report compiled to PDF with its BibTeX, where Exa Agent returns text and JSON.
• Systematic reviews: research questions, screening of up to 3,000 papers, extraction, a PRISMA flow diagram and a report. Web search APIs do not offer screening, PRISMA or citation export.
• Extraction tables from full text: AI-filled columns for datasets, methods, sample sizes or outcomes, which you can then filter, sort and combine.
• Live streaming and follow-ups: like Exa Agent it streams progress; it also streams the answer as it is written, and follow-ups continue in the same thread.

When to use which

• Use Exa when you need control over retrieved URLs, domain filters and page content for your own pipeline, general web research with schema-defined JSON output, or its larger publication index.
• Use the SciSpace API when the deliverable is a screened, extracted and cited literature report delivered as files.

If scholarly recall matters, test it yourself. In Elicit's own benchmark (August 2026) on 5,486 biomedical questions, Exa found 40.3% of relevant papers in its top 50, against 60.3% for Elicit's scholarly search. SciSpace was not part of that test.

What it doesn't do

• Not a web search API: for news, documentation or general web pages, keep Exa or another web search API.
• No paywalled papers: it reads your own PDFs plus 50M+ open-access full texts, so some papers will only have abstract-level evidence.
• Only the sources listed above: it does not query Semantic Scholar, Scopus, Web of Science, Lens or other sources not named here.

The SciSpace API is in early access; it isn't open yet. At launch it's open to anyone: register on SciSpace and start, with no sales call. Every account gets $5 in free API credits, enough to try both APIs on real questions. The API is separate from SciSpace consumer plans; no consumer plan includes API usage.

Frequently asked questions

Last checked

October 21, 2026. Exa facts were checked against official documentation. SciSpace API details reflect its early-access status. No authenticated comparative benchmark was performed.

Sources

1. Exa API pricing
2. Exa pricing documentation
3. Exa billing and rate limits
4. Exa Search API reference
5. Exa Answer API reference
6. Exa Agent API: Create a run
7. Exa Agent quickstart
8. Exa research publications documentation

注意转载授权 · 阅读原文