
Best Deep Research APIs in 2026: 8 Compared
On this page
- Most of 2025’s deep research endpoints have been retired. OpenAI shut down
o3-deep-research, Perplexity retired legacy Sonar Chat Completions, and Exa retired/research. Any comparison written before mid-2026 points at deprecated APIs. - Pricing now splits cleanly into two models: fixed per-task tiers (Valyu, Parallel, Linkup and Exa fixed-effort) versus metered token-plus-search billing (OpenAI Responses, Google Gemini Interactions, Perplexity Agent and Tavily dynamic credits).
- For high-volume company enrichment into strict JSON with flat unit economics, start with Valyu DeepResearch or Parallel (Valyu when filings, financials, or specialist records matter; Parallel for pure web rows). For general web synthesis inside an existing lab contract, OpenAI, Perplexity or Google fit cleanly.
- Choose Valyu DeepResearch when the brief needs specialist primary datasets alongside the web (SEC filings, PubMed, arXiv, ClinicalTrials.gov, patents or licensed Wiley journals), passage-level citation locations, 2026+ point-in-time date pinning, or real
xlsx,pptx,docxandpdfdeliverables from the same API call.
Deep research began in 2025 as a button inside consumer chat apps: type a question, wait fifteen minutes, and copy a 3,000-word essay into a document. Over the past year it moved behind an API boundary. Engineering teams now wire asynchronous research directly into private-equity diligence pipelines, hedge-fund forecasting and backtesting loops, pharma trial trackers, R&D patent checks and internal agent workflows where no human sits in a chat window.
At the same time, almost every provider rebuilt its API surface. OpenAI retired o3-deep-research and o4-mini-deep-research in July 2026, folding multi-step research into the Responses API. Perplexity ended legacy Sonar Chat Completions in September 2026 and repriced its Agent API in November. Exa retired /research in April 2026 in favour of Deep Search and the Exa Agent API. Google opened Gemini Deep Research on the Interactions API, while Parallel, Linkup, Tavily and Valyu shipped dedicated research endpoints with sharply different pricing, source coverage and output shapes.
Across the eight production APIs available in November 2026, the gap in answer quality on ordinary web questions has largely closed. The differences that matter in production show up in four operational places: what happens to your bill when an autonomous loop searches eighty pages, whether the agent reads primary filings and trials or web commentary about them, whether a citation points to a URL or an exact passage, and whether the output arrives as raw markdown or a finished file a team can open and act on.
- 2025 endpoints retired in 2026
- 3
- OpenAI o3-deep-research, Perplexity Sonar, Exa /research
- Published cost range per task
- $0.01–$21+
- Parallel base / Exa minimal to multi-agent ultra tiers
- Spend swing on uncapped token loops
- 5–10×
- Autonomous reasoning + search billing on complex briefs
- Production APIs compared
- 8
- Verified against live pricing and docs, November 2026
Which deep research API should you choose?
Every provider below can browse the web and write a cited summary. Pick based on the constraint that actually binds your workflow: primary data coverage, unit economics, output format or existing cloud procurement.
| If your product needs to… | Start with | Why |
|---|---|---|
| Run private equity diligence, hedge fund forecasting, pharma/biotech or R&D research from primary sources | Valyu DeepResearch | Queries SEC filings, PubMed, clinical trials, patents and Wiley journals alongside the web in one task, with passage-level citations and 2026+ point-in-time date pinning for clean backtests. |
| Return Excel models, slide decks, Word memos or branded PDFs straight from an API call | Valyu DeepResearch | Generates up to 10 files (xlsx, pptx, docx, pdf, csv) plus Brand Kit PDFs from the same research run. |
| Enrich thousands of companies into strict JSON with flat unit economics | Valyu DeepResearch or Parallel Task API | Both offer fixed per-run pricing with zero token math and failed runs unbilled. Choose Valyu when company, finance, or science enrichment needs primary filings, patents, or research records alongside web data; choose Parallel for pure web-only rows. |
| Build prospect lists, enrich contacts or combine fixed tiers with a web index | Exa Agent API | Fixed tiers ($0.012 to $1.00) or capped metered mode, built-in email/phone enrichment and Exa’s own web crawl. |
| Run fast, low-cost web synthesis with built-in step budgets | Perplexity Agent API | November 2026 presets put high-effort web research around $0.12–$0.14 per request with automatic step limits. |
| Stay inside an existing OpenAI Responses stack with custom tools | OpenAI Responses API | Uses the same background Responses primitives, reasoning effort controls and enterprise billing you already run. |
| Combine Google Search grounding with large uploaded document sets on GCP | Google Gemini Interactions API | Pairs Gemini’s long context and File Search with Google Search grounding inside Google Cloud. |
| Stream progress events into a UI or buy through AWS Marketplace | Tavily Research | Native SSE progress streaming, selectable academic citation styles (APA, MLA, Chicago) and AWS Marketplace distribution. |
| Get flat-rate web and AFP news research with simple S/M/L/XL tiers | Linkup Deep Research | Flat $0.25 to $2.50 per run with answer, investigate and research shapes; failed or empty runs are unbilled. |
The 8 deep research APIs compared
The table below reflects published list prices, active endpoints and documented capabilities verified in November 2026. Pay-as-you-go rates are shown unless noted; monthly plans and enterprise commitments lower effective per-task prices on credit-based providers.
| Provider & active endpoint | Billing model | Published price / task | Sources | Outputs & deliverables | Citations & execution |
|---|---|---|---|---|---|
| Valyu DeepResearch POST /v1/deepresearch/tasks | Fixed credits per mode + optional tools/files (failed tasks unbilled) | instant $0.05 · fast $0.16 · standard $0.63 · heavy $3.13 · max $15.63 (PAYG) | Web + specialist indexes: SEC filings, PubMed, arXiv, bioRxiv/medRxiv, ChemRxiv, ClinicalTrials.gov, USPTO/EPO patents, FRED/BLS, licensed Wiley HLS | Markdown, PDF, JSON Schema, charts, Brand Kits, plus up to 10 deliverables (xlsx, pptx, docx, pdf, csv) | Numbered citations + passage-level sources[].locations (excerpt & page where captured); async polling, webhooks, batch, HITL |
| Parallel Task API Task API / Responses | Flat USD per successful run (failed runs unbilled) | lite $0.005 · base $0.01 · core $0.025–$0.05 · pro $0.10 · ultra $0.30 to ultra8x $2.40 | Web index and web retrieval | Markdown report or schema-enforced JSON | Field-level citations and confidence basis on structured outputs; async up to 2 hours |
| Exa Agent API Agent API / Deep Search | Fixed-effort tiers or metered ACUs ($0.10/ACU + $0.005/search) | Fixed: $0.012 (minimal) to $1.00 (xhigh). Metered: capped at $5 (auto) or $20 (ultra) | Exa web index + optional contact enrichment & Exa Connect providers | Markdown report or structured JSON | Inline URL citations and source snippets; async polling & webhooks |
| OpenAI Responses API + web_search | Model input/output tokens + $10 per 1,000 web searches | Variable: typically $0.50–$10+ depending on model (e.g. GPT-6 Sol $2/$10 per 1M) and step count | OpenAI web_search + user-supplied tools/files | Text, markdown or Structured Outputs JSON | Inline URL annotations; async background mode + streaming |
| Perplexity Agent API | Underlying model tokens + per-tool fees ($2.50/1k web, $5/1k finance/people) | Published benchmark averages: high ~$0.12–$0.14 · xhigh ~$0.22–$0.78 · ultra ~$21.71 | Perplexity web search, URL fetch, finance/people tools, code sandbox | Text, markdown or JSON; preset step budgets | Inline URL citations; sync or async presets |
| Google Gemini Interactions API (preview) | Gemini tokens ($2/$12 per 1M ≤200k) + $14/1k Google searches (after 5k free/mo) | Google estimates: ~$1.00–$3.00 (Standard) · ~$3.00–$7.00 (Max) | Google Search grounding + File Search / url_context | Markdown report (preview) | Grounding metadata & source links; async background=True |
| Linkup POST /research | Flat USD per run by reasoning depth (failed/empty unbilled) | S $0.25 · M $0.50 · L $1.50 · XL $2.50 (sync Deep Search $0.05–$0.055) | Web search + AFP dispatches for authorised accounts | Natural-language report (answer, investigate, research) or JSON Schema | Inline citations & sources array; async 1–20 min |
| Tavily POST /research | Dynamic credits per call ($0.008/cr PAYG; down to $0.005/cr on Growth) | mini $0.032–$0.88 · pro $0.12–$2.00 (PAYG) | Web search + base64 .txt/.md/.json file grounding | Synthesised report or JSON output_schema | Numbered, MLA, APA or Chicago citations; async polling or SSE streaming |
Comparison of entry-tier and deep-tier per-task prices in US dollars across Valyu, Parallel, Exa, Linkup, Tavily, Perplexity and Google Gemini as of November 2026.
- Valyu DeepResearchinstant vs standard (PAYG)
- Valyu DeepResearchstandard vs heavy (PAYG)
- Parallel Task APIpro vs ultra8x (fixed)
- Exa Agent APImedium vs xhigh (fixed)
- Linkup Deep ResearchS vs XL (fixed)
- Tavily Researchmini floor vs pro ceiling (PAYG)
- Perplexity Agent APIhigh vs xhigh (published avg)
- Google GeminiStandard vs Max (est. midpoint)
The four places deep research breaks in production
Most evaluations fail because teams test three prompts in a notebook and compare the writing style of the output. Writing style is the easiest part to change. Once a research API is wired into a live product or a recurring internal batch, four operational fault lines decide whether it stays in production.
1. Uncapped token billing turns hard briefs into budget surprises
An autonomous research loop decides at runtime how many searches to run and how many long documents to pull into context. When an API bills every intermediate reasoning token and search step — as OpenAI’s Responses API, Google’s Interactions API and Perplexity’s Agent API do — two prompts that look identical to a user can differ by 5× to 10× in cost. A query that hits a dead end and retries across eighty pages quietly burns through your margin on that customer seat.
Fixed-price tiers solve this at the contract layer. Parallel, Linkup, Exa’s fixed-effort tiers and Valyu DeepResearch price the task up front and absorb the variance of searches and page reads inside the run. On Valyu, creating a task reserves its base credit price before work begins, returns 402 cleanly if an API key or monthly spend cap is hit, and never charges for failed tasks. For batch runs across 300 portfolio companies or fixed-price SaaS features, predictable unit economics matter more than shaving a few cents off an average benchmark run.
2. Web search returns commentary about primary records, not the records
Ask a web-only deep research API to pull a target company’s debt from its latest 10-K, check whether a Phase 3 drug trial met its main endpoint, or review the claims on a competitor’s patent. Most of what the agent retrieves will be secondary commentary: press releases, blog recaps and scraped aggregator pages that paraphrase the original document.
For casual market scans that is fine. For private equity deal teams, hedge fund and forecasting desks, biotech and pharma teams or corporate R&D groups, paraphrased commentary introduces errors you only catch when someone opens the primary document. That is the core split between general web research APIs and specialist data infrastructure: whether the research agent can query SEC filings, academic papers and preprints, ClinicalTrials.gov records, USPTO and EPO patent text, FRED and BLS economic series, or licensed publisher content such as Wiley’s health and life sciences journals directly alongside the open web.
3. A URL is not an audit trail
Every API in this list returns citations. Very few make those citations fast to audit. If a report claims a medical-device manufacturer’s gross margin fell 140 basis points and links to the root URL of a 140-page Form 10-K, an analyst still has to open the PDF, search for the right table and reconcile the period. Multiply that by forty numbers in a memo and the time savings vanish.
Two patterns are emerging to close that gap. Parallel attaches field-level citation basis and confidence to structured JSON outputs so downstream software knows which URL supports which field. Valyu DeepResearch goes a step further on long-form reports and document sources by returning sources[].locations — an optional field containing the exact supporting excerpt, page number and link fragment where a stored passage is available — so your application can display the sentence from the filing beside the claim it supports.
4. Raw markdown still leaves a human copying into Excel
Consumer chat apps such as ChatGPT, Claude and Gemini can already generate spreadsheets and slide decks inside their own interfaces. Inside an API pipeline, however, almost every deep research endpoint still returns either a markdown string or a JSON object. That works when your software writes directly to a database, but if the end recipient is a deal team, a consulting partner or a medical-affairs director, someone still spends an hour moving markdown tables into a workbook or rebuilding slides.
Valyu DeepResearch treats deliverables as first-class API outputs: you can request up to 10 files per task across xlsx, pptx, docx, pdf and csv (the first deliverable is included in the task price, with additional files at 16 credits each and Excel/PowerPoint/Word requiring code_execution), and apply your organisation’s styling with Brand Kits. The research agent produces the underlying analysis and the finished artefact in one asynchronous run.
Each deep research API in one paragraph
Valyu DeepResearch. Built for high-stakes financial services (private equity diligence, hedge-fund forecasting and backtesting), life sciences and pharma, and corporate R&D where primary sources and finished deliverables matter. Five fixed-price modes cover the full latency ladder — from instant (8 credits, short cited answer in under 30 seconds, replacing Valyu’s deprecated Answer API) through fast (25 credits), standard (100 credits), heavy (500 credits) and max (2,500 credits), with PAYG credits priced at $10 per 1,600 credits, monthly subscriptions lowering per-credit rates by 8% to 25%, and new work-email accounts starting with 2,000 free credits each month. It combines web search with specialist indexes (SEC filings, PubMed, arXiv, bioRxiv, medRxiv, ChemRxiv, ClinicalTrials.gov, USPTO/EPO patents, FRED/BLS and licensed Wiley journals), returns passage-level sources[].locations and up to 10 xlsx/pptx/docx/pdf/csv files, and supports point-in-time reference_date runs with historical_cache_strict: "only" for 2026+ snapshots so backtests drop pages lacking an in-window capture rather than falling back to a live crawl.
Parallel Task API. Alongside Valyu, one of the strongest flat-rate options for high-volume structured company enrichment — specifically when the workload is pure web data. Parallel’s processor ladder runs from lite ($0.005) and pro ($0.10) up to ultra8x ($2.40) with zero token math, no charges on failed runs, and schema-enforced JSON outputs backed by field-level citations. If your enrichment pipeline only needs open-web signals across thousands of rows at a fixed dollar cost, Parallel is hard to beat; when enrichment or diligence touches SEC filings, financials, clinical trials, patents, or scientific literature, pair or replace it with Valyu DeepResearch.
Exa Agent API. Replaced Exa’s legacy /research endpoint in April 2026 and pairs nicely with Exa’s own web index. Teams can choose predictable fixed-effort tiers (minimal at $0.012 up to xhigh at $1.00 per run) or metered Agent Compute Units ($0.10/ACU plus $0.005 per search, with hard defaults of $5 on auto and $20 on ultra), plus optional email/phone enrichment and Exa Connect partner feeds. Strong for go-to-market research, list building and web synthesis; outputs are markdown or JSON rather than spreadsheets or decks.
OpenAI Responses API. After shutting down o3-deep-research and o4-mini-deep-research in July 2026, OpenAI now runs deep research through background Responses calls with high reasoning effort and the web_search tool ($10 per 1,000 searches plus full input/output token rates, such as $2/$10 per 1M tokens on GPT-6 / GPT-6.1 Sol or $4/$20 on GPT-5.6 Sol). The upside is tight control: you pick the exact frontier model, mix in your own custom tools and stay on one OpenAI invoice. The downside is spend variance — long autonomous runs can swing from under $1 to $10+ unless you explicitly cap tool calls and output tokens — and retrieval stays limited to web search plus tools you wire yourself.
Perplexity Agent API. Replaced legacy Sonar Chat Completions in September 2026 and received a major price reduction alongside a new multi-agent ultra tier in November 2026. Pricing combines model token rates with per-tool fees ($2.50/1k standard web searches, $5/1k finance or people searches, $0.03 per 20-minute sandbox), packaged into managed presets whose published benchmark averages run ~$0.12–$0.14 on high (up to 40 steps), ~$0.22–$0.78 on xhigh (up to 300 steps) and ~$21.71 on parallel multi-agent ultra. Built-in step budgets keep routine web synthesis fast and inexpensive, though you are still paying variable token bills and working with web-centric text/JSON outputs.
Google Gemini Interactions API. Exposes Google’s Deep Research Agent (deep-research-preview-04-2026 and Max) asynchronously via background=True. Billing combines Gemini token rates ($2 input / $12 output per 1M tokens up to 200k context, with long-context surcharges above) and Google Search grounding ($14 per 1,000 searches after 5,000 free monthly calls). Google’s own per-task estimates sit at $1–$3 for Standard (~80 searches) and $3–$7 for Max (~160 searches). It fits teams already standardised on Google Cloud who want to blend Google Search with large uploaded document collections via File Search, provided preview status and variable token spend are acceptable.
Linkup Deep Research. Offers a straightforward flat-rate async /research API across four tiers — S ($0.25), M ($0.50), L ($1.50) and XL ($2.50) — alongside a synchronous depth="deep" search mode at $0.05–$0.055 per call, with failed and empty queries unbilled and SOC 2 Type II across plans. Outputs support structured JSON or three text formats (answer, investigate, research), plus AFP news dispatches for authorised accounts. A clean, predictable choice for web and news investigations without specialist scientific or filing indexes.
Tavily Research. A dynamic credit-based /research endpoint priced at 4–110 credits on mini and 15–250 credits on pro (billed at $0.008 per credit PAYG, dropping to $0.005 per credit on Growth). Standout ergonomic features include real-time Server-Sent Events streaming of tool calls and progress, base64 text/markdown/JSON grounding attachments, structured output_schema and selectable citation formats (numbered, mla, apa, chicago). It is especially easy to adopt if your stack already uses Tavily or buys via AWS Marketplace.
Build your own search loop or call a deep research API?
Before committing to any deep research endpoint, ask whether you should orchestrate your own agent on top of a search API instead. As we noted when looking at why leading investment funds build in-house research platforms, the boundary between building and buying is about where your proprietary edge actually lives: in the research orchestration logic or in the underlying data and execution.
Keep a custom Search + LLM loop when…
You own the planner, prompt state and every tool call yourself, and call Search / Contents APIs underneath.
- Your product’s moat is a proprietary multi-step scoring rubric or private internal-document join.
- Full control over every tool call, prompt and state transition
- Latency must stay under five seconds for interactive chat or single-hop retrieval.
- Direct Search API calls avoid async task overhead
- You already maintain retries, context compaction, deduplication and eval harnesses in-house.
- Incremental cost is raw model tokens plus per-result search fees
Call a managed Deep Research API when…
You delegate multi-hop planning, source reading, synthesis and file creation to a single async task.
- A brief requires 10 to 50+ follow-up queries, page reads, calculations and cross-checks.
- Skip months of harness engineering, rate-limit queuing and context-rot fixes
- You need predictable unit economics per report or batch job without token blowups.
- Fixed per-mode pricing caps cost per accepted output
- End users expect audit-ready citations, charts or formatted xlsx/pptx/docx/pdf files.
- Deliverables and passage locations ship out of the box
How to evaluate deep research APIs on your own workload
Public leaderboards tell you which systems are in the same tier; they will not tell you which one survives your domain’s edge cases. Run a narrow, reproducible bake-off before signing an annual contract:
- Assemble 15 to 20 real briefs with known ground truth. Pull actual questions your users or analysts ran last month, and record the primary-source answer (the filing line item, the trial NCT number, the patent claim) before running any API.
- Pin equivalent budget tiers, not marketing labels. Compare a $0.50–$0.63 task against a $0.50–$0.63 task (for example Valyu
standard, Exahigh, LinkupMand Parallelultra2x) rather than comparing a 5-cent fast lookup against a $5 deep run. - Audit primary-source recall and citation checkability. Measure how many claims cite the primary document directly versus a secondary blog post, and time how long it takes a human reviewer to verify five randomly chosen numbers.
- Track p50 and p95 cost and wall-clock time. On variable-billing APIs, the p95 cost is what breaks unit economics. Record timeouts, failed tasks and whether failed runs were billed.
- Inspect the handoff into your downstream format. If your users need a JSON object, spreadsheet or slide deck, include the cost and failure rate of the post-processing step you would otherwise have to build yourself.
To include Valyu DeepResearch in that test set, a single task call looks like this. Swap included_sources between presets such as finance, medical, patent and web (or specific dataset IDs like valyu/valyu-sec-filings) to test specialist coverage, and read sources[].locations from the status endpoint when you want exact passage excerpts.
from valyu import Valyuvalyu = Valyu() # reads VALYU_API_KEY from the environmenttask = valyu.deepresearch.create(query=("Compare revenue growth, EBITDA margin and net debt for five publicly ""traded US medtech companies using their latest annual filings."),mode="standard",search={"included_sources": ["finance", "web"],},tools={"code_execution": {"enabled": True},},deliverables=[{"type": "xlsx","description": "A Summary tab with one row per company and a Sources tab.",},],)result = valyu.deepresearch.wait(task.deepresearch_id)if result.status == "completed":print(f"Total cost: ${result.cost:.2f} ({result.cost_credits} credits)")for d in result.deliverables or []:print(d.type, d.status, d.url)
Frequently asked questions
What is the best deep research API in 2026?
It depends on your data sources, pricing model and output format. For specialist knowledge work that combines the web with SEC filings, PubMed, arXiv, clinical trials, patents or licensed Wiley journals — or when you need passage-level source locations and real xlsx, pptx, docx or pdf files from the API — Valyu DeepResearch is the strongest fit. For high-volume web enrichment into strict JSON at flat per-run rates, Parallel Task API and Exa Agent API are strong defaults. For general web synthesis inside an existing frontier-lab contract, OpenAI’s Responses API, Perplexity’s Agent API and Google’s Gemini Interactions API cover most needs.
How much does a deep research API cost per task?
In November 2026, entry-level research calls cost roughly $0.01 to $0.25 per task (Exa minimal $0.012, Valyu instant $0.05 PAYG, Parallel pro $0.10, Perplexity high ~$0.12–$0.14, Linkup S $0.25). Mid-tier multi-step reports typically run $0.50 to $3.15 (Linkup M $0.50, Exa high $0.50, Valyu standard $0.63 and heavy $3.13 PAYG, Parallel ultra8x $2.40, Google Standard ~$1–$3). Exhaustive multi-agent tiers reach $15 to $21+ per task (Valyu max $15.63 PAYG, Exa metered ultra up to $20, Perplexity ultra ~$21.71 on published benchmarks).
What happened to OpenAI o3-deep-research, Perplexity Sonar and Exa Research?
All three 2025 endpoints were retired in 2026. Exa retired /research on April 1, 2026 in favour of Deep Search and the Exa Agent API. OpenAI shut down o3-deep-research and o4-mini-deep-research on July 23, 2026, migrating workloads to the Responses API with background mode and web_search. Perplexity ended legacy Sonar Chat Completions on September 27, 2026, replacing sonar-deep-research with managed effort presets on its Agent API.
Which deep research APIs have fixed per-task pricing?
Valyu DeepResearch, Parallel Task API, Linkup Deep Research and Exa’s fixed-effort Agent tiers all offer fixed per-task pricing known before the run starts, and none of them charge for failed tasks. OpenAI Responses, Google Gemini Interactions, Perplexity Agent API, Exa’s metered mode and Tavily Research bill dynamically based on tokens, credits or compute units consumed during the run.
Which deep research API returns Excel, PowerPoint, Word or PDF files?
Valyu DeepResearch natively generates downloadable xlsx, pptx, docx, pdf and csv deliverables alongside the report — up to 10 files per task in every mode except instant — and supports Brand Kits for styled PDFs. Most other deep research APIs return markdown text or schema-validated JSON, leaving spreadsheet and slide generation to your own application code.
Should I build my own deep research loop with a search API instead?
Build your own loop on top of a search API when your core differentiation is custom orchestration logic, private-database joins or sub-five-second latency. Use a managed deep research API when briefs require 10 to 50+ iterative searches, long-document synthesis, predictable per-report pricing or finished deliverables — building and maintaining a reliable multi-hop research harness usually costs more engineering time than the API margin.
Test on your hardest 15 briefs
Run 15 real questions from your team against Valyu DeepResearch with the primary datasets your work actually depends on, and inspect the sources, passage excerpts, deliverables and cost_detail side by side with your current setup. Create a free account (2,000 free credits every month with a work email, no card required), read the DeepResearch documentation, check pricing and plan concurrency or compare raw retrieval providers in our 2026 Web Search APIs guide.
If you're an agent: Set up Valyu and run your first deep research task.
If you're a human: Create an account and get an API key.
MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.
More from the blog





Valyu Add
Join 12,000+ professionals and knowledge workers.
Valyu Add is a free weekly research briefing for builders, investors and operators. Every issue is sourced, cited and verified with Valyu DeepResearch.
