Best AI Research Tools: ChatGPT, Gemini, Claude, Perplexity, Elicit and Three More Scored

A research agent turns a question into a plan, runs dozens of searches, reads what it finds and returns a cited report in minutes. Every major AI lab now ships one, and a group of scholarly tools does the same over paper databases. This guide scores eight of them on how deep they go, how checkable their citations are, what they can search and how easy they are to get.

What counts as an AI research agent?

A research agent decides its own next step. It breaks a question into sub-questions, searches each, reads full sources, notices what is missing and searches again before it writes. AI search answers one query from one round of retrieval; a research agent runs many rounds. OpenAI says a deep research task takes 5 to 30 minutes (OpenAI).

The market splits three ways. General agents (ChatGPT, Gemini, Perplexity) search the open web. Work-data agents (Claude Research, Microsoft 365 Copilot Researcher) add email, documents and meetings. Scholarly agents (Elicit, Consensus, Undermind) search scientific literature. If you publish content and want these agents to cite you, that is a different problem; see our agentic search guide.

How do the research agents score?

Four criteria, 1 to 5 each, from vendor documentation, help centers, announcements and pricing pages through October 31, 2026. These are editorial assessments of published capabilities, not benchmark results.

• Depth: how thorough a task can be: iteration, steering, higher-effort modes.
• Citations: how verifiable the output is: inline citations, source restriction, scholarly corpora, citation checking.
• Reach: what it can search: web, uploaded files, work apps, MCP servers, paper databases.
• Access: free tier, published prices, plan availability and an API.

The research agents

Gemini Deep Research

Google launched Deep Research in the Gemini app in December 2024 and has upgraded the model behind it with each Gemini release. On April 21, 2026 it announced two API agents on Gemini 3.1 Pro: Deep Research for speed and Deep Research Max for depth, both able to query MCP servers and private data next to the web (VentureBeat). Google reported 93.3% for Max on its DeepSearchQA benchmark. In the API, both are in preview through the Interactions API with collaborative planning, File Search and chart generation (Google AI for Developers). In the Gemini app, Deep Research is listed on every plan from free to AI Ultra, counted against compute-based usage limits (9to5Google). Best fit: anyone who wants the widest free access, and developers building research into a product. Full Gemini Deep Research review.

ChatGPT deep research

OpenAI launched deep research on February 2, 2025. A February 2026 update moved it to a GPT-5.2-based model, added the option to limit research to specific sites, connections to apps and MCP servers, and a better report view (Wikipedia). Users can interrupt a run to steer it or add sources. The API offers o3-deep-research and o4-mini-deep-research with MCP and webhooks. OpenAI no longer publishes fixed task counts for personal plans: access is labelled limited on Free and Go, expanded on Plus ($20 a month) and maximum on Pro (from $100 a month), with Business at $25 per seat (ToolColumn, citing OpenAI, Oct 11, 2026). Best fit: the strongest all-round web researcher for people already on ChatGPT. Full ChatGPT deep research review.

Claude Research

Anthropic added Research to Claude in April 2025 together with a Google Workspace integration (Anthropic). It runs searches that build on each other across the web, Gmail, Google Calendar, Google Docs and any custom connectors enabled in the chat. It is available on Pro, Max, Team and Enterprise, runs on Claude Opus 5.5 with Opus 4.8 as fallback whatever model the chat uses, and draws from the same usage limits as normal chats, faster (Claude Help Center). On mobile it searches only the web. Best fit: teams on Google Workspace who want research that includes their own mail and documents. Full Claude Research review.

Elicit

Elicit searches more than 138 million papers and 500,000 clinical trials. It added Research Agent workflows in December 2025 and a standalone Research Agent in August 2026, which now also reaches patents and safety reports; Routines (September 2026) re-run research on a schedule, and the systematic review workflow supports PRISMA 2020 (Elicit blog). An API and MCP server let other agents use it. Basic is free; Plus is $11 per user per month billed annually, with Pro, Scale and Enterprise above it (Elicit pricing). Elicit raised a $22 million Series A in February 2025. Best fit: literature reviews and evidence work that must stand up to scrutiny. Full Elicit review.

Perplexity Deep Research

Perplexity launched Deep Research on February 14, 2025 with free daily queries (Engadget). Its Advanced Deep Research update added clarifying questions, follow-ups during a run, a code sandbox, document uploads and reports that stream into an editable file; Max subscribers get Claude Opus 4.6 Thinking as the model (Perplexity Help Center). Pro was $20 a month and Max $200 a month in April 2026, and the Sonar Deep Research API is billed by tokens plus search fees (Finout). Perplexity changed its usage limits during 2026, so check the current caps. Best fit: fast, citation-first reports for people who already use Perplexity for search. Full Perplexity Deep Research review.

Microsoft 365 Copilot Researcher

Researcher is the Microsoft 365 Copilot agent for multistep research across the web and the files, emails, meetings and chats a user can access, with cited reports that can be turned into documents or presentations (Microsoft Support). In March 2026 Microsoft added a Critique mode in which a GPT model drafts and Claude reviews the draft for accuracy, completeness and citation integrity (The Letter Two). It needs a Microsoft 365 Premium or Pro subscription, or a Microsoft 365 Copilot licence at work, and an admin must enable it. Best fit: Microsoft 365 organizations whose answers sit in SharePoint, Outlook and Teams. Full Copilot Researcher review.

Consensus

Consensus is a search engine over scientific literature that answers from papers, with a Consensus Meter that shows how many studies support or contradict a yes/no question and a Deep Search mode that writes a structured report in a few minutes. It raised $30 million in May 2026 and says 2.5 million researchers start work with it each month (Signalbase). A free tier exists; a third-party guide lists Pro at $15 a month with 15 Deep Searches (Pillitteri), but Consensus's own pricing page did not show prices when we checked. Best fit: fast evidence checks on health, science and policy questions. Full Consensus review.

Undermind

Undermind (Y Combinator, Summer 2024) is a deep search agent for scientific papers. A search reviews thousands of papers, reads full texts of open-access PDFs and stops after about 2.5 minutes on average; results can be monitored for new publications (Undermind). Its own benchmark claims 73% recall against 29 to 39% for frontier-model agents, which has not been independently checked. It connects to ChatGPT and Claude through MCP. Free plan; Pro is $16 a month and Team $15 per person a month, billed annually (Undermind pricing). Best fit: researchers who need to find every relevant paper, not a summary of the top ten. Full Undermind review.

Also worth knowing

• NotebookLM Deep Research: Google's notebook tool runs its own Deep Research with daily caps by plan. See NotebookLM alternatives.
• You.com ARI: a business research agent launched in February 2025; we could not confirm its 2026 plans, so it is not scored.
• Build your own: teams that want a research agent inside their product can start from the Gemini, OpenAI or Perplexity research APIs or an orchestration framework; see agent orchestration frameworks.

How should you choose a research agent?

1. Start from where the answer lives. Open web: Gemini, ChatGPT or Perplexity. Company data: Claude Research or Copilot Researcher. Papers: Elicit, Consensus or Undermind.
2. Check data controls before features. Which connectors can it read, can admins scope them, and is your data used for training on the plan you buy?
3. Test citations on your own questions. Run five questions you already know the answer to and open every citation in two reports.
4. Model usage, not seats. Research tasks are rationed by plan and burn compute fast; heavy users hit caps first.
5. Keep a person on anything that leaves the team. Numbers and quotes get traced to a primary source before they go into a deck, a filing or a client deliverable.

Related questions

• How do research agents work, and how should a company govern them? (partner site CTAIO)
• ChatGPT, Gemini or Perplexity: which deep research should you use?
• Elicit, Consensus or Undermind for scientific literature?
• How do you get your pages cited by research agents?
• How did reinforcement learning shape how deep research browses?

Sources

• OpenAI, Introducing deep research (Feb 2, 2025, updated Feb 10, 2026); Wikipedia, ChatGPT Deep Research; ToolColumn, ChatGPT deep research limits by plan (Oct 11, 2026); The Decoder, OpenAI expands API access to deep research models (Jun 26, 2025).
• VentureBeat, Google's new Deep Research and Deep Research Max agents (Apr 21, 2026); Google AI for Developers, Gemini Deep Research agent; 9to5Google, Google AI Plus, Pro and Ultra features (May 25, 2026).
• Anthropic, Claude takes research to new places (Apr 15, 2025); Claude Help Center, Use research on Claude.
• Engadget, Perplexity has its own Deep Research tool now too (Feb 15, 2025); Perplexity Help Center, What's new in Advanced Deep Research; Finout, Perplexity pricing in 2026 (Apr 27, 2026).
• Microsoft Support, Get started with Researcher; The Letter Two, Researcher upgrade with multi-model intelligence (Mar 30, 2026).
• Elicit, Blog and Pricing.
• Signalbase, Consensus funding (May 2026); Pasquale Pillitteri, Consensus AI guide 2026 (third-party).
• Undermind, Home and Pricing; Hunted/Product Hunt, Undermind (YC S24) launch (Aug 20, 2024).
• Venkit et al., DeepTRACE, arXiv 2509.04499 (Sep 2, 2025).

Related reading on our partner site: agentic research governance tiers and hallucinations at system scale on CTAIO.

What is the best AI research agent in 2026?

On published capabilities, Gemini Deep Research and ChatGPT deep research lead for general web research, Claude Research and Microsoft 365 Copilot Researcher lead for research across company email, documents and meetings, and Elicit leads for scientific literature. The best pick depends on where the answer lives: the open web, your work data, or peer-reviewed papers.

What is an AI research agent?

An AI research agent takes a question, plans a series of searches, reads the sources it finds, searches again to fill gaps and writes a report with citations. A single task usually takes minutes rather than seconds. Most vendors call the feature deep research; Anthropic calls it Research and Microsoft calls it Researcher.

Is there a free deep research tool?

Yes. Gemini Deep Research is available on the free Gemini plan within standard usage limits, ChatGPT offers limited deep research on Free and Go, Perplexity has a free tier, and Elicit, Consensus and Undermind all have free plans. Free tiers ration the number of research tasks.

Can I trust the citations in a deep research report?

Check them. An independent audit (DeepTRACE, September 2025) found citation accuracy between 40 and 80 percent across deep research systems, and OpenAI says its own deep research can hallucinate facts. Scholarly agents such as Elicit, Consensus and Undermind cite papers rather than web pages, which makes checking easier but not optional.

Which AI research agent can search my company data?

Claude Research can search Gmail, Google Calendar, Google Docs and other enabled connectors. Microsoft 365 Copilot Researcher searches files, emails, meetings and chats the user can access. ChatGPT deep research connects to apps and MCP servers, and Google's Deep Research agents in the Gemini API accept MCP servers and File Search over your own documents.