Best Deep Research APIs: Pricing, Speed & Benchmarks

Best deep research APIs at a glance

You need more than search. The API has to plan searches, read and cross-reference sources, then return a cited report you can use programmatically. The problem is that vendor benchmark claims often conflict because they use different tests, model variants, and publishers.

The benchmark column is contradictory on purpose. The next section explains why these scores are not directly comparable.

What is a deep research API?

A deep research API runs an iterative research loop rather than a single retrieval call. It typically breaks a question into sub-questions, searches for relevant sources, reads those sources, cross-references claims, and synthesizes the findings into a cited answer.

That is the main difference from a web search API. A search API usually returns links, snippets, rankings, or extracted page content. A deep research API goes further: it decides what to investigate next, combines evidence from multiple sources, resolves conflicting information where possible, and returns a report with citations.

This extra work explains the runtime. A normal search request can complete in milliseconds or seconds. Deep research can take minutes or, for heavier research modes, hours because the agent may run many searches, open many documents, inspect PDFs, revisit earlier assumptions, and generate a longer final response.

You should not use a deep research API for every query. It is the wrong tool for simple fact lookups, short Q&A, or any task where one search already gives you the answer. These calls are slower and more expensive because they execute multiple retrieval and reasoning steps behind a single request.

Use one when the quality of the answer depends on multi-source investigation, not just finding the right page.

Why the benchmark scores don't agree

Perplexity Sonar Deep Research scores 26.9% on DeepSearchQA when Parallel measures it and 70.5% on DRACO when Valyu measures it. That gap is large enough that you should not treat either number as a universal measure of research quality.

The first problem is that these are not neutral league tables. Vendors choose which benchmark to publish, which configuration to run, and which competitor product or tier to compare against. Even the table above does not always compare identical product configurations. Parallel is represented by Ultra2x on DeepSearchQA and Ultra8x on DRACO. OpenAI is represented by GPT 5.2 Pro on one benchmark and Deep Research o3 on the other. Exa is tested as Research Pro in one case and Deep Reasoning in the other.

That makes a headline ranking much less useful than it first appears.

The benchmarks also test different things. DeepSearchQA, created by Google DeepMind, contains 900 multi-step tasks across 17 fields. DRACO, Perplexity's own benchmark, contains 100 long-form tasks across 10 professional domains. A model can perform differently when the task distribution, grading method, expected report format, and research depth change.

There is another problem: the published DeepSearchQA claims do not even agree with each other.

Parallel reports its own Ultra2x system at 72.6%, Gemini Deep Research at 64.3%, OpenAI GPT 5.2 Pro at 61%, Exa Research Pro at 27%, and Perplexity Sonar Deep Research at 26.9%. Perplexity separately claims state of the art on DeepSearchQA and Scale AI's ResearchRubrics. You.com markets its Research API as "#1 on DeepSearchQA."

Those statements cannot all function as a single, stable ranking. They may come from different configurations, evaluation procedures, system versions, or testing conditions. The practical conclusion is simpler: when the vendor publishing the benchmark also wins the benchmark, treat the result as marketing evidence, not an independent ranking.

This problem gets harder with public evaluations. A 2026 arXiv paper measures search-time contamination in deep research agents, where agents can encounter public benchmark material or related evaluation information during web research. That can inflate measured performance without producing the same gain on unseen production queries.

Published scores are still useful. They can tell you which systems are worth testing and expose large differences under a specific evaluation setup. They should not decide the vendor for you.

For a production choice, build your own evaluation set of 20 to 50 queries drawn from the workload you actually expect to run. Keep the questions, source requirements, output format, and grading criteria fixed across providers. Then compare citation quality, factual accuracy, source coverage, report usefulness, runtime, and cost under the configurations you would actually deploy.

Treat every published benchmark score as a directional claim, not a final ranking.

How we grouped these APIs

We do not publish a 1-to-8 ranking. The benchmark data in the previous section does not support one, and these products do not all solve the same problem.

Instead, we group them by the job they perform.

• ‍Tier A, full agentic research: you send a question and get back a cited report. This includes Parallel, You.com, Valyu, and Perplexity.‍
• Tier B, research primitives: you get search, extraction, and retrieval endpoints that you compose into your own research loop. This includes Exa, Tavily, and Firecrawl.‍

Best full agentic research APIs in 2026

Parallel Task API

What it's for. Parallel Task API is built for full agentic research, from short answers to long-form reports and structured outputs.

How it works. It runs on Parallel's proprietary web index of billions of pages. The system uses token-efficient ranking to select relevant material, then performs live crawling when needed, including JS-heavy pages and PDFs. This gives it both indexed retrieval and fresh page access within the same research flow.

Pricing. Pricing is split across nine processor tiers. Published prices include Lite at $5 per 1,000 queries, Pro at $100 per 1,000, Ultra at $300 per 1,000, Ultra2x at $600 per 1,000, and Ultra8x at $2,400 per 1,000. Pricing for the remaining tiers is not published in the data provided.

Runtime. Not published.

Output. The API can return short summaries, full reports, or structured data. Citation format and audit-trail details are not published in the data provided.

Benchmark claims. Parallel published its own DeepSearchQA results: Ultra2x scored 72.6%, Ultra scored 68.5%, and Pro scored 62%. A separate DRACO evaluation published by Valyu scored Parallel Ultra8x at 51.2%.

Best for. Teams that want a single research API spanning lightweight summaries, deeper reports, and structured outputs, with access to both a large proprietary index and live crawling.

Watch out for. The headline DeepSearchQA scores are self-published by Parallel. A third-party result from Valyu places Ultra8x at 51.2% on DRACO, so you should not treat the vendor's DeepSearchQA figures as an independent ranking. Runtime is also not published, and the available pricing data does not include all nine processor tiers.

You.com Research API

What it's for. You.com Research API is built for full agentic research that turns a question into a cited report after searching and cross-checking multiple sources.

How it works. The system first plans a research strategy, then runs multiple searches, reads the returned material, and cross-references sources before generating the final answer. Its ARI system can read up to 400 sources, while ARI Enterprise extends that to 500+ sources across the web, internal data, and premium data.

Pricing. Lite costs $6.50 per 1,000 queries, Standard $50 per 1,000, Deep $100 per 1,000, and Exhaustive $300 per 1,000. A Frontier tier is also available, but its price is not published in the data provided.

Runtime. ARI completes in under five minutes.

Output. The Research API returns Markdown with inline citations. ARI can also generate cited PDF reports with interactive visualizations. A separate Finance Research API is available for finance-specific research workloads.

Benchmark claims. You.com markets its Research API as "#1 on DeepSearchQA", a claim published by You.com itself. Valyu separately scored You.com at 52.9% on DRACO.

Best for. Teams that want a research system capable of reading hundreds of sources and returning cited Markdown or presentation-ready PDF reports, especially when internal or premium sources need to be included through ARI Enterprise.

Watch out for. The "#1 on DeepSearchQA" claim is self-published, so it should not be treated as an independent ranking. The Frontier tier price is not published. ARI Enterprise also has broader source access than the standard ARI configuration, so make sure you compare the tier you would actually deploy.

Linkup

What it's for. Linkup is a French AI search provider covering fast web retrieval through to asynchronous multi-step research.

How it works. Linkup exposes three main endpoints: Fetch for retrieving specific pages, Search for sub-second sourced answers with full-text snippets, and Research for asynchronous multi-step queries. You can tune the index, source selection, and freshness parameters depending on the workload.

Pricing. Standard Search costs $0.005 per request, or $5 per 1,000 queries. Deep Search costs $0.05 per request, or $50 per 1,000, exactly 10x the standard rate. Research is priced by reasoning depth at $0.25, $0.50, $1.50, or $2.50 per run, equivalent to $250, $500, $1,500, or $2,500 per 1,000 queries. Linkup also offers 4,000 free queries per month or a $20 monthly credit. Linkup's own blog quotes "€5 / 1,000" while other sources quote $5, so confirm the billing currency before modelling costs.

Runtime. /search runs in under one second. Runtime for the asynchronous /research endpoint is not published. Linkup reports 99.9% uptime with SLA backing.

Output. Search returns sourced and cited answers with full-text snippets. Research supports multi-step queries, while the available data does not specify additional report file formats or audit-trail exports.

Benchmark claims. Linkup reports a 92% F-score on Verified SimpleQA for its sub-second search API and 61% on SealQA-0 for /research, claiming first place on both respective comparisons. Both results were published by Linkup. Unlike the other headline scores in this article, Linkup also publishes a public reproducible evaluation harness for Verified SimpleQA at github.com/LinkupPlatform/eval-simpleQA.

Best for. Teams that want both low-latency search and deeper asynchronous research, with explicit controls over sources and freshness. Linkup integrates with the OpenAI SDK, LangChain, and CrewAI, and is available through Eden AI alongside Firecrawl.

Watch out for. The benchmark results are still vendor-published, but Linkup is the only provider in this article that ships a public reproducible eval harness. Every other headline benchmark score covered here is not independently reproducible from the material provided. Also confirm whether standard Search billing is denominated in euros or dollars.

Valyu DeepResearch

What it's for. Valyu DeepResearch is built for long-form research that needs access to specialist sources beyond the open web.

How it works. It runs multi-step research across public web content plus specialised and proprietary datasets. Its source coverage includes SEC filings, PubMed, bioRxiv, medRxiv, ClinicalTrials.gov, USPTO and global patents, FDA documents, FRED, BLS, World Bank, IMF, real-time markets, academic publishing, and legal records.

Pricing. Valyu offers four tiers: Fast, Standard, Heavy, and Max. The vendor publishes figures of $100, $500, $2,500, and "$15 per request", but the units are internally inconsistent on its pricing page. You should verify the billing unit directly before modelling cost per 1,000 queries.

Runtime. Fast takes 1-8 minutes, Standard 5-20 minutes, Heavy up to 90 minutes, and Max up to 3 hours.

Output. Valyu can return Markdown, PDF, DOCX, XLSX, and PPTX. It can generate spreadsheet formulas and render charts from primary sources. Inline citations trace back to source passages, numerical claims are verified before inclusion, and webpage screenshots can be embedded as an audit trail.

Benchmark claims. Valyu reports 72.7% on DRACO, 74.5% on ScholarQA-CS, and 4.56/5 on ScholarQA-Multi. All three results were published by Valyu.

Best for. Research workflows that depend on primary financial, scientific, regulatory, patent, legal, or market data, especially when the final deliverable needs to be exported as a document, spreadsheet, or presentation.

Watch out for. The published pricing units are inconsistent, so cost cannot be converted reliably into a clean per-1,000-query figure from the available data. The headline benchmark results are also self-published by Valyu.

Perplexity Sonar Deep Research

What it's for. Perplexity Sonar Deep Research is built for multi-step web research that returns a synthesized answer with detailed citation metadata.

How it works. The system performs repeated retrieval and synthesis rather than a single search call. It works within a 128,000-token context window and adjusts research depth per request, which also affects billing.

Pricing. Perplexity does not publish a fixed cost per 1,000 deep research queries. Pricing combines $2 per million input tokens, $8 per million output tokens, and $5 per 1,000 searches, plus citation tokens. As of 2026, citation charges apply only to Deep Research. The per-request fee also varies with search depth, so total cost per query is difficult to predict in advance.

Runtime. Most runs complete in under three minutes. Based on the data in this comparison, it is the fastest turnkey deep research option covered here.

Output. The API returns multi-step retrieval and synthesis results with deep citation metadata. Other output formats are not published in the data provided.

Benchmark claims. Perplexity claims state-of-the-art performance on DeepSearchQA and Scale AI's ResearchRubrics. Those claims are published by Perplexity. Valyu scored Perplexity Deep Research at 70.5% on DRACO. Parallel scored Perplexity Sonar Deep Research at 26.9% on DeepSearchQA.

Best for. Teams that want a turnkey research API with sub-three-minute runtime for most runs and detailed citation metadata.

Watch out for. Cost forecasting is the main issue. Billing combines input tokens, output tokens, searches, citation tokens, and a variable per-request fee tied to search depth. The benchmark picture is also inconsistent: Perplexity claims state of the art on DeepSearchQA, while Parallel measured Sonar Deep Research at 26.9% on that same benchmark.

Best research primitives you compose yourself

Exa

What it's for. Exa is a research primitive for semantic discovery, deep search, and answer generation rather than a pure long-form report engine.

How it works. Exa uses embeddings-native semantic search to retrieve pages by meaning rather than relying only on keyword matching. Its product surface includes standard search, deep search, deep-reasoning search, an Answer endpoint, and an Agent API for more agentic workflows.

Pricing. Search costs $7 per 1,000 queries, Deep Search $12 per 1,000, Deep-Reasoning Search $15 per 1,000, and the Answer endpoint $5 per 1,000. Page summaries add $1 per 1,000. The base price includes up to 10 results. The Agent API uses agentComputeUnits instead of fixed per-query pricing. Default "auto" effort caps at $5 per run, while beta max caps at $20. Fixed effort levels bill a flat amount per request.

Runtime. Not published.

Output. Exa provides search results, deep-search results, answers, and optional page summaries. Other report formats, inline citation behavior, and audit-trail details are not published in the data provided.

Benchmark claims. Parallel scored Exa Research Pro at 27% on DeepSearchQA. Valyu scored Exa Deep Reasoning at 25.9% on DRACO. Both results were published by competitors, not by Exa.

Best for. Teams building their own research loop that need semantic retrieval and want separate search, reasoning, answer, and agent components.

Watch out for. The benchmark scores need context. Exa is primarily a discovery tool being tested on report-writing tasks, so low synthesis scores do not mean the underlying search product performs poorly. The Agent API also uses a different billing model from the standard endpoints, which makes direct per-1,000-query cost comparisons less clean.

Tavily

What it's for. Tavily is a research primitive for teams that want search, extraction, crawling, mapping, and a dedicated /research endpoint in the same API surface.

How it works. Tavily exposes separate endpoints for search, extract, crawl, map, and /research. The /research endpoint launched in 2025. The data provided does not specify the internal research mechanism, so details such as planning strategy, indexing method, or model architecture are not published here.

Pricing. Tavily prices usage in credits rather than a fixed per-query research fee. Growth costs $0.005 per credit, with the listed plan at $500 per month for 100K credits. Pay-as-you-go costs $0.008 per credit. A free tier is available. A clean cost per 1,000 /research queries is not published because the data provided does not state how many credits one research request consumes.

Runtime. Not published for /research. Tavily reports a 180ms median latency at the platform level, but that figure should not be treated as the runtime of the research endpoint.

Output. Tavily exposes search, extraction, research, crawl, and map outputs. Specific /research report formats, citation behavior, and audit-trail details are not published in the data provided.

Benchmark claims. Not published in the data provided.

Best for. Teams building their own research workflow that want search and extraction primitives alongside a dedicated research endpoint, with published rate limits of 100 RPM on free accounts and 1,000 RPM on paid accounts.

Watch out for. Tavily does not publish /research runtime or a fixed per-1,000-query cost in the data provided. Its reported 180ms median latency applies at the platform level, not specifically to deep research. Tavily reports 300M+ monthly requests, 99.99% uptime, and 2M+ developers. Nebius acquired the company in February 2026 for $275M, with the total potentially reaching $400M through performance milestones.

Firecrawl

What it's for. Firecrawl is best treated as a web research primitive for search, extraction, and agent-driven batch workflows, not as a standalone deep research endpoint.

How it works. The old v1 deep research endpoint is deprecated. Firecrawl now points users toward its Search API, which combines web search and content extraction in one call and returns LLM-ready Markdown, and its Agent endpoint, which supports batch processing across many queries. Firecrawl also points users to its open-source Firesearch project.

Pricing. The legacy deep research endpoint bills 1 credit per URL analyzed. A fixed cost per 1,000 research queries is not published in the data provided because total cost depends on how many URLs each request analyzes.

Runtime. The legacy endpoint supports a timeLimit between 30 and 300 seconds. It runs asynchronously with progress tracking. Runtime for the replacement Search API and Agent endpoint is not published in the data provided.

Output. The Search API returns LLM-ready Markdown after combining search and extraction. The legacy research endpoint supports asynchronous progress tracking. Other report formats, inline citation behavior, and audit-trail details are not published in the data provided.

Benchmark claims. Not published in the data provided.

Best for. Teams that want web search and extraction components they can compose into their own research workflow. Firecrawl is also the web provider available through Eden AI.

Watch out for. The v1 deep research endpoint is deprecated, and Firecrawl's own documentation gives contradictory dates. Deprecation was announced on 2 February 2026, while the documentation states that the endpoint remains active until 30 June 2025. Do not rely on either date without checking the current endpoint status. The replacement products also use a different workflow, so the legacy endpoint's pricing and runtime should not be assumed to apply to Search or Agent.

Google Gemini Deep Research

What it's for. Gemini Deep Research is relevant if you already use Google's AI stack, but it is not cleanly packaged as a standalone deep research API.

How it works. Deep Research is available in the consumer Gemini app across all subscription tiers, including free. Programmatic access exists in Google's broader stack, but a clean standalone Vertex AI endpoint dedicated to deep research could not be confirmed from the data provided. Deep Think is separate and runs through the AI Ultra subscription rather than as a billed API model.

Pricing. A fixed Deep Research API price is not published. Grounded search on Vertex costs roughly $35 per 1,000 requests. That figure should not be treated as a direct price for Gemini Deep Research itself.

Runtime. Not published.

Output. Deep Research produces research-style synthesized answers in the Gemini product. API-specific output formats, citation metadata, structured export options, and audit-trail behavior are not published in the data provided.

Benchmark claims. Parallel scored Gemini Deep Research at 64.3% on DeepSearchQA and estimated its cost at roughly $3,680 per 1,000 queries. Valyu scored Gemini Deep Research at 59.0% on DRACO. The first result was published by Parallel, the second by Valyu.

Best for. Teams already working inside Google's AI ecosystem that can accept some ambiguity around how deep research is exposed programmatically.

Watch out for. The main issue is product packaging. Deep Research exists, and programmatic access exists around Gemini, but a clean standalone Vertex AI deep research endpoint could not be confirmed. Do not use the roughly $35 per 1,000 grounded-search figure as if it were a published Deep Research API price. The benchmark cost estimate of roughly $3,680 per 1,000 comes from Parallel, not Google.

What deep research actually costs per 1,000 queries

Pricing is only directly comparable when the vendor charges per query. For token, credit, or URL-based billing, the data available here does not include enough workload assumptions to produce a defensible normalized figure. Those rows are marked accordingly rather than filled with invented estimates.

Why per-token research pricing is hard to forecast

Perplexity is the clearest example. A Deep Research request can incur input-token charges, output-token charges, citation-token charges, and $5 per 1,000 searches. The per-request fee also changes with search depth.

That means an estimate needs assumptions for average input tokens, output tokens, citation tokens, and searches per query. APPENDIX A provides none of those averages, so any dollar figure per 1,000 queries would be invented. Research workloads make this especially variable because two user questions can trigger very different search depths.

Parallel and You.com take a different approach. Both sell explicit effort tiers. With Parallel, you can move from $5/1K Lite to $2,400/1K Ultra8x. You.com ranges from $6.50/1K Lite through $300/1K Exhaustive, plus a Frontier tier with no published price. For these APIs, cost is a dial you control, not one fixed rate.

Valyu needs separate caution. Its page lists Fast, Standard, Heavy, and Max alongside $100, $500, $2,500, and "$15 per request", but the published units are internally inconsistent. Confirm the billing basis with Valyu before building a cost model.

The bigger lesson is that higher effort does not translate linearly into benchmark gains. Parallel's own DeepSearchQA ladder shows Pro at $100/1K and 62%, Ultra at $300/1K and 68.5%, and Ultra2x at $600/1K and 72.6%, all figures published by Parallel. Moving from Pro to Ultra2x costs six times more for roughly ten percentage points on that benchmark. Tier names and price gaps can therefore overstate the practical performance gap. Test the cheapest tier that meets your workload before moving up.

How to choose a deep research API

1. Do you need a finished report or structured data your code consumes?

If you want to send one question and receive a completed research output, start with the Tier A providers. Parallel can return short summaries, full reports, and structured data. You.com returns Markdown with inline citations and can produce cited PDFs, while Valyu supports Markdown, PDF, DOCX, XLSX, and PPTX. If structured machine-consumable output matters most, Parallel is the clearest fit from the data here.

2. What is your latency budget?

Runtime varies widely. Perplexity says most Deep Research runs complete in under three minutes, You.com ARI completes in under five minutes, and Valyu ranges from 1-8 minutes on Fast to as much as three hours on Max. If you need a turnkey research result quickly, Perplexity is the best fit in this comparison. If longer jobs are acceptable in exchange for deeper research modes, Valyu gives you the widest published runtime range.

3. Do you need specialised sources such as SEC filings, PubMed, patents, FDA documents, or economic data?

This is the easiest decision path. Valyu is the only provider in this article with native coverage explicitly listed for SEC filings, PubMed, bioRxiv, medRxiv, ClinicalTrials.gov, patents, FDA documents, FRED, BLS, World Bank, IMF, legal records, and real-time markets. If those sources are central to the workload, start there.

4. Do you need predictable cost, or is variable spend acceptable?

For easier forecasting, Parallel and You.com publish fixed per-1,000-query effort tiers. You can choose a research depth and know the base query cost before running it. Perplexity is harder to forecast because it combines input tokens, output tokens, citation tokens, searches, and a variable per-request fee tied to search depth. If predictability matters, start with Parallel or You.com.

5. Do you want a turnkey agent, or search and extraction primitives you compose yourself?

For a finished agentic workflow, choose among Parallel, You.com, Valyu, or Perplexity. If you want to own the research loop and assemble retrieval, extraction, reasoning, and synthesis yourself, look at Exa, Tavily, or Firecrawl instead. Exa focuses on semantic retrieval, Tavily exposes search, extract, crawl, map, and research endpoints, and Firecrawl combines search with content extraction and agent-style batch processing.

Do not commit based on the published benchmark tables alone. As the benchmark section shows, vendor rankings conflict sharply, so validate your shortlist on your own representative queries before choosing a production route.

Implementation patterns

Getting structured output instead of prose

Keep the request simple, then extract only the fields your application needs from the returned JSON. You can validate those fields later in your application if needed.

import requests

url = "https://api.edenai.run/v3/universal-ai/"
headers = {
    "Authorization": "Bearer <your-api-key>",
    "Content-Type": "application/json"
}
payload = {
    "model": "web/research_async/firecrawl",
    "input": {
        "prompt": "Research <your topic>"
    },
    "show_original_response": False
}

response = requests.post(url, headers=headers, json=payload)
data = response.json()

print(data)

‍

For production use, keep only the fields you need instead of storing the complete response.

Handling long-running jobs

Deep research jobs can take minutes, so submit the request asynchronously rather than keeping one HTTP connection open. Save the returned job ID, then use Eden AI's async result endpoint to check when the result is ready.

import requests

url = "https://api.edenai.run/v3/universal-ai/"
headers = {
    "Authorization": "Bearer <your-api-key>",
    "Content-Type": "application/json"
}
payload = {
    "model": "web/research_async/firecrawl",
    "input": {
        "prompt": "<your research query>"
    },
    "show_original_response": False
}

response = requests.post(url, headers=headers, json=payload)
data = response.json()

print(data)
# Save the returned job ID and poll Eden AI's async result endpoint.

‍

Check Eden AI's documentation for the exact result endpoint and returned job ID field.

Capping cost per run

Do not send provider parameters that the endpoint does not support. The simplest approach is to apply your budget rules before the research request and control how often expensive research jobs can run.

import requests

url = "https://api.edenai.run/v3/universal-ai/"
headers = {
    "Authorization": "Bearer <your-api-key>",
    "Content-Type": "application/json"
}
payload = {
    "model": "web/research_async/firecrawl",
    "input": {
        "prompt": "<your research query>"
    },
    "show_original_response": False
}

# Check your application budget before sending the request.

response = requests.post(url, headers=headers, json=payload)
print(response.json())

‍

For providers with effort or tier controls, select the appropriate tier before submitting the query and enforce the spending limit in your own application.

Running deep research through Eden AI

If you are testing more than one research provider, the main cost is often integration switching rather than the API call itself. Eden AI reduces that switching cost by exposing providers through a single API gateway with one key, one response format, and consolidated billing.

Its Universal AI endpoint covers web/search, web/scraping, and web/crawl_async, backed by Firecrawl and Linkup. Switching between Firecrawl and Linkup only requires changing one segment of the model string, while the rest of the request stays the same.

For deep research specifically, Linkup's /research endpoint is the option available through Eden AI today.

Eden AI charges a 5.5% platform fee on top of provider usage. That tradeoff makes most sense when you expect to test, compare, or switch providers rather than integrate each one separately.

FAQs - Best Deep Research APIs: Pricing, Speed & Benchmarks