Best AI research agents in 2026: deep research tools compared

Published by 3Gen Internet Corporation, maker of OpenAssistant, Skej and AGNT. Comparisons use public documentation, without hands-on testing.

For the best AI research agents, start with the dedicated research products. Elicit leads for scientific literature and evidence work, and OpenAI’s deep research in ChatGPT leads for broad, cited web reports, with Perplexity, Gemini and Claude close behind. Almost every one of them hides its real usage limit behind a plan label, a billing toggle or a default view, so read the limits section of each entry before you pay. If you want research that arrives in your email, text or Slack and feeds into scheduling and follow-ups, see the short OpenAssistant section below.

Options at a glance

We ordered the products against the site rubric: whether the documented capabilities cover the research job, the vendor-published evidence and its limitations, price for what you get, where it works, and privacy, security and admin controls.

• Elicit ranks first because its pages describe the most complete research workflow for evidence work: a stated 138 million paper index, claim-level links to source passages, uploads, and paid tiers for systematic reviews. It is narrower than a general web tool, and its training policy is stated only for Enterprise.
• OpenAI deep research ranks second for documented controls that include: an editable plan, uploaded files, connected apps, site restrictions, citations on every output, and admin role-based access. Its current limits are not published as numbers, which makes its limits harder to compare.
• Perplexity ranks third for a fast stated run time, uploaded documents and a low annual entry price, offset by training on by default for Free, Pro and Max and no numeric limit for Pro.
• Gemini ranks fourth for a free tier that includes Deep Research and Google Workspace sources, held back by no stated citation format, no published request caps and a human-review warning in its privacy text.
• Claude Research ranks fifth: clear model and connector documentation and citations, but paid plans only, shared limits and no stated file or export details.
• Consensus ranks sixth because only its pricing page was available, so its deep review mechanics are not documented in what we could read.
• Manus Wide Research ranks seventh because it targets a different job (the same question across 10 or more items), and its citation behavior and privacy policy were not available.

1. Elicit

“Every claim links back to the exact supporting passage, table, or figure” (Elicit Research Agent page).

Elicit’s Research Agent searches the literature, extracts data and builds reports, tables and slides. The vendor says it searches over 138 million academic papers and 545,000 clinical trials, plus patents, regulatory sources and the open web, and accepts uploads in BibTeX, EndNote, Mendeley, PDF and Zotero formats (Elicit). The pricing page says Pro adds systematic review screening of 5,000 papers, 20 data columns and API access (Elicit pricing).

Price: The pricing page has an Academic view and an Industry view. Confirm the relevant Academic or Industry pricing view and current plan with the linked source (Elicit pricing).

Watch out for:

• The page loads in the Industry view by default, so a reader sees the higher prices first. Annual Pro is $49 per month in Industry against $39 in Academic, a $10 difference, and annual Scale is $229 against $89, a $140 difference.
• Usage multiples change with the audience too. Academic Pro is labelled 5x usage, while Industry Pro is labelled “standard usage”. Numbers behind “standard usage” are not stated.
• Basic is free with “limited usage” for the Research Agent and Research Reports, with no figure given.
• “No training on your data by default” is listed under Enterprise only. The training policy for the other plans is not stated on the pricing page, and the privacy page could not be read.

2. OpenAI deep research in ChatGPT

“reasoning, researching, and synthesizing information into a documented report” (OpenAI help center).

ChatGPT proposes a research plan you can edit before it starts, lets you watch progress and interrupt, and works from the public web, uploaded files and connected apps. The help article says every output includes citations or source links, with a table of contents, a sources-used section and downloads as Markdown, Word or PDF (OpenAI help center). A February 2026 update to the launch post adds MCP and app connections and a trusted-site restriction (OpenAI launch post). Apps are read-only for research, and Enterprise and Edu admins control access through role-based access control.

Price: The pricing page lists Free at $0, Go at $8 per month, Plus at $20 per month and Pro from $100 per month with a choice of three usage tiers (ChatGPT pricing). Business seats are $20 per month billed annually or $25 monthly, and the premium seat is $100 annual or $125 monthly (ChatGPT pricing, Business and Enterprise tab). Go may include ads.

Watch out for:

• Current limits are not published as numbers. The help article says usage varies by plan and points to an in-product counter, and fixed allowances reset every 30 days from first use. The pricing grid says only Limited for Free and Go, and Yes for Plus and Pro (ChatGPT pricing).
• The only numbers on OpenAI’s pages are historical. An April 2025 update to the launch post says Plus, Team, Enterprise and Edu get 25 queries a month, Pro 250 and Free 5, then a lightweight version takes over. The same post’s body says Pro gets up to 100 a month. Treat every number as historical, not as a current allowance.
• The launch post gives a run time of 5 to 30 minutes. The current help article gives none.
• OpenAI’s own limitations list says the tool “can sometimes hallucinate facts” and may struggle to separate authoritative sources from rumors (OpenAI launch post).
• The benchmark figure OpenAI gives, 26.6% on Humanity’s Last Exam, is a vendor claim from the February 2025 launch.
• For Free, Go, Plus and Pro, the pricing page lists content as used for training with an opt-out. Business and Enterprise default to no training.

3. Perplexity Research

“performs dozens of searches, reads hundreds of sources” (Perplexity Help Center).

Perplexity calls the feature Research mode in help articles, Advanced Deep Research in its update notes and “Deep research” on its pricing page. The vendor says it refines its plan as it learns, asks clarifying questions for broad queries, processes uploaded documents and exports to PDF, a document or a Perplexity Page (What is Research mode, what’s new). You cannot pick a model for Research.

Price: Pro is $17 per month billed annually and Max is $167 per month billed annually (Perplexity Pro page). A help article gives Max as $200 monthly or $2,000 annually, which matches $167 a month (Perplexity Max). Enterprise Pro starts at $40 per seat per month or $400 per seat per year, and Education Pro is $10 per month with verification (plan comparison).

Watch out for:

• Pro’s monthly-billing price is not shown on the pages read, so the $17 figure is annual only.
• The plan table gives Free 1 Research query per month, Enterprise Pro 50 a month and Enterprise Max 500 a month. Pro and Max show only “Monthly limits (average use)” and “(advanced use)” with no number. Another article says Free users get “limited access”, and the pricing page does not repeat the 1 per month figure.
• Limits reset on a rolling basis. The vendor says renewing or reaching your billing date does not reset them.
• The Research article gives two run times: “under 3 minutes” and “around 4 to 5 minutes”.
• Free, Pro and Max have AI training on by default with an opt-out. The vendor says previously collected training data cannot be deleted after you opt out. Enterprise data is never used for training (Perplexity data collection).
• No accuracy warning appeared on the Research articles read.

4. Google Gemini Deep Research

“an agentic feature in Gemini that can automatically browse up to hundreds of websites” (Gemini Deep Research overview).

Gemini plans, searches, reasons and reports. The help article says a report usually takes about 5 to 10 minutes, you can edit the plan, and you can leave and get notified. Sources include Google Search by default, Gmail, Drive, uploaded files and NotebookLM notebooks, and reports export to Docs (Gemini help).

Price: Free is $0 and includes Deep Research. Google AI Plus is $4.99 per month, AI Pro $19.99 per month, and AI Ultra from $99.99 per month, or $199.99 for the 20x tier (Gemini subscriptions). A Business edition starts at $21 per seat per month. Annual prices are not shown.

Watch out for:

• No request cap is published. The help article names daily and concurrent limits without numbers. The limits page says compute-based limits refresh every 5 hours up to a weekly limit, rates Deep Research as high cost, says it may be unavailable on the free tier at busy times, and says limits may change without notice (Gemini limits).
• Plus is 2x the standard limit, Pro 4x, and Ultra 5x or 20x of Pro.
• Charts, diagrams and simulators are Ultra-only and unavailable if you include Google Workspace sources.
• The overview and help article do not describe a citation format.
• The privacy page says human reviewers review some data, asks you not to enter confidential information, and says reviewed chats can be kept for up to three years (Gemini Apps privacy).

5. Claude Research

“runs multiple searches that build on each other” (Claude help center).

Claude decides what to look into next, researches the web and connected tools, and Anthropic’s launch post says answers come with “easy-to-check citations” (Anthropic). The comparison guide says a run takes five or more tool calls over 1 to 3 minutes (Claude guide). Web search must be on. Research runs on Claude Opus 5.5 with a fallback to Opus 4.8, whatever model your chat uses.

Price: Free is $0 with no Research. Pro is $20 billed monthly or $17 per month on annual billing ($200 billed up front). Max starts at $100 per month, with 5x or 20x usage. Team standard seats are $20 per seat annual or $25 monthly, and Enterprise is $20 per seat per month plus usage at API rates with a 20-seat minimum (Claude pricing).

Watch out for:

• There is no separate research counter. Research shares your normal limits and “can use up your limits faster”.
• Connector tools run automatically, “without asking for approval”, and the mobile app does not use connectors and gives no error. Local MCP servers in Claude Desktop cannot be used.
• The Max 20x price is not printed as its own figure in the page text read, and Max is monthly billing only.
• File uploads, export formats and accuracy cautions are not stated on the articles read.

6. Consensus

“Consensus is the AI-powered academic search engine” (Consensus pricing).

Consensus claims more than 400 million peer-reviewed papers. Only its pricing page was available, so how a Deep review works, how it cites, and its privacy terms are not stated in the material we reviewed.

Price: The page defaults to the annual view (Consensus pricing).

Watch out for:

• The headline price is the annual one. Pro costs $20 on monthly billing, and the annual saving is $96 a year ($240 against $144). Deep costs $65 monthly, and the annual saving is $240 a year ($780 against $540).
• Free search is basic, “without any AI analysis”. Free includes 10 Pro messages a month.
• Pro includes 500 API and MCP uses and Deep 2,000, with overage at $0.05 per use up to 1,000 a month on Pro and 10,000 on Deep.
• Students, faculty and clinicians can get up to 40% off.

7. Manus Wide Research

“deploys hundreds of independent agents that work in parallel” (Manus docs).

Wide Research splits one request across many sub-agents, each with its own context, and assembles the answers into a table, report or dataset. The docs say it typically takes minutes for 50 to 100 items and name 250 items as the largest size they cover. They advise against it for fewer than 10 items or a single deep dive, where regular agent mode fits better. It is a list-and-compare tool, not a literature tool. “Uniform quality at any scale” is a vendor claim (Manus Wide Research).

Price: The pricing page shows three paid cards without plan names (Manus pricing).

Each card also lists 300 daily refresh credits.

Watch out for:

• Credits are consumed by task complexity, and the credit cost of a Wide Research run is not stated (Manus plans). Plan credits reset monthly, and purchased add-on credits do not expire.
• The docs list a Free plan, but the pricing page showed no free card.
• Citation behavior, file handling and the privacy policy are not stated on the pages read.

OpenAssistant: research that arrives where you already work

OpenAssistant is a general work assistant, not a research product. It has its own name, email address, phone number and calendar, and takes requests by email, text, calls and Slack. Its pages describe looking into companies, people and markets and bringing back “a brief instead of a list of links”, and a help-center example asks it to research a company’s business model, funding, leadership and competitors, with a report and spreadsheet (OpenAssistant executive assistant page, Ways to work). The same page says it chases replies and reminds people of what they owe you.

Dedicated deep-research products are built for the research job. Choose Elicit, ChatGPT or another entry above for a literature review or a long cited report. Choose OpenAssistant when the brief should land in a Slack DM, a text or an email thread and the next step is a meeting or a follow-up.

• Price: The Starter plan is free, with 1,000 credits at signup plus 100 bonus credits a day once used up, up to 1,000 a month. Pro is $72 per user per month, or $60 annual, and Team is $120 per user per month, or $100 annual, at the 5,000-credit tier (OpenAssistant pricing; per-user pricing confirmed by 3Gen). The pricing notes say deep research uses hundreds of credits where a simple task uses 10 to 50, and plan credits do not roll over (OpenAssistant pricing notes).
• Controls: Autonomy is set per action type as Manual, Confirmation Window or Autonomous. A Confirmation Window is not Manual approval (OpenAssistant autonomy settings). The trust page states a SOC 2 Type II audit and that it never trains on your data (OpenAssistant trust page).
• Watch out for: It does not join group texts or group WhatsApp chats, and does not work in Slack Connect channels with outside companies (Ways to work and Slack help). How its reports cite sources is not stated on the pages read.

What to ask each vendor

• What is my exact monthly allowance of deep research runs on this plan, and what happens when I hit it?
• Which billing view (monthly or annual, academic or industry) does the price I was shown assume?
• If a page I research contains hidden instructions (prompt injection), what stops the agent from acting on them, especially through connectors that run without approval?
• Are my prompts, uploads and reports used for model training on this plan, and can I switch that off for data already collected?
• Does every claim in the report link to a source passage, and can I export those links?
• Which connected apps can the agent read, and can an admin turn them off?

For more on risks, see are AI agents safe, and for how credits and seats are billed, see AI agent pricing explained. For wider picks by job, see best AI agents in 2026.

What to keep for humans

• Checking the sources behind any claim you will act on. OpenAI’s own limitations list admits wrong inferences and weak confidence signals.
• Deciding what a finding means for your project, patients, clients or budget.
• Choosing what confidential material goes into a tool whose plan allows training or human review.
• Approving anything that leaves your organization, such as an email or a message sent from your account.
• Judging whether a source is authoritative.

Related reading

• Best AI browser and computer-use agents in 2026, compared: Nine browser and computer-use agents compared on documented controls, where the browser runs, approvals and prices.
• OpenAssistant vs Manus: two ways to hand work to an AI agent: Channels, approvals, prices, free plans, security statements and limits for OpenAssistant and Manus.
• Best free AI agents in 2026: what you get at no cost: True free plans, trials and expiring credits, with the stated limits and what each free tier lacks.
• How to choose an AI agent: 10 questions to ask before you buy: Ten questions on job fit, channels, approvals, price, data and exit, with a job-to-product table.
• Best AI agents in 2026: top picks by job, with prices: Our recommendations for each job (general work, coding, browser use, research, sales, support, enterprise and builders), with vendor prices.

Frequently asked questions

What is the best AI research agent in 2026?

From public information, Elicit is the strongest pick for scientific literature and evidence work, and OpenAI deep research in ChatGPT is the strongest for broad cited web reports. The ranking is an editorial judgment, not a measured result.

Are there free AI research agents?

Several have free tiers with small limits. Elicit Basic has limited Research Agent use, Consensus Free allows up to 3 Deep reviews a month, Perplexity’s plan table gives Free 1 Research query a month, and Gemini’s free plan includes Deep Research. Claude Research is not on the Free plan, and the Manus docs list a Free plan that the pricing page did not show.

How many deep research reports do I get each month?

Mostly the vendors do not say. ChatGPT, Perplexity Pro and Max, and Gemini publish no current numbers. Consensus publishes 3, 15 and 200 Deep reviews a month. Perplexity’s Enterprise Pro and Max tables give 50 and 500 a month. Old OpenAI numbers (25, 250 and 5) date from April 2025.

Can I trust the citations in a deep research report?

Citations help you check a claim, but they do not prove it. OpenAI says its tool can hallucinate and mix up rumor with authority, and Elicit says each claim links back to a source passage. Open the sources before you rely on anything.

Can OpenAssistant do deep research?

Its pages describe research briefs on companies, people and markets, and note that deep research uses hundreds of credits. It is a general assistant, so a dedicated tool is the better fit for a literature review.

Try OpenAssistant

If you want research that shows up in your email, texts or Slack and turns into meetings and follow-ups, try OpenAssistant on the free Starter plan. See the OpenAssistant pricing page for plan details.