Products, laws and numbers here change fast. Checked October 2026.
In plain words
A normal chatbot answer comes in seconds, from what the model already knows plus perhaps one quick search. Deep research is slower on purpose. You ask a bigger question, such as "Which electric scooter suits a hilly city on a budget?", and the AI goes off to work.
It makes a plan, splits the question into parts, runs dozens of web searches, opens and reads pages, follows up on what it finds, and then writes an organised report with links to its sources. It usually takes a few minutes, sometimes half an hour.
Google, OpenAI, Perplexity and Anthropic all launched versions between December 2024 and April 2025. It's useful for a first pass on a topic. It can still pick weak sources or misread them, so treat the report as a well-organised starting point and click the citations that matter.
Everyday example
Sending a junior to the library for the afternoon
You ask a new intern: "Find out which coaching institutes in our city have the best results for engineering entrances." They make a list of what to check, read websites, reviews and news, note where each fact came from, and hand you a two-page summary by evening. Deep research does that afternoon's work in minutes.
Where the example breaks: An intern can phone an institute or notice that a review site looks paid for. The AI only sees what is on the open web or in files you connect, and can repeat a confident claim from a poor source.
How it actually works
Deep research is an agent specialised for one job: research. It loops through searching, reading and deciding what to look up next.
Plan
A reasoning model breaks your question into sub-questions and a search plan. Gemini shows you the plan to approve or edit before starting.
Search and read in a loop
It runs searches with tool use, opens promising pages and PDFs, extracts what matters and starts new searches based on what it learned. Some systems run several sub-agents in parallel, each on one angle.
Keep the context manageable
Reading hundreds of pages would overflow the model's context window, so findings are summarised as it goes, a form of context engineering.
Write and cite
The model writes a structured report and attaches citations to claims. Anthropic, for example, uses a separate step just to match each claim to its source. Good practice for you: open the citations behind the key numbers.
- Launches
- Google Gemini (11 December 2024), OpenAI ChatGPT (2 February 2025), Perplexity (14 February 2025), Anthropic Claude Research (15 April 2025)
- How long
- OpenAI's version browses for 5 to 30 minutes; Perplexity said 2 to 4 minutes for its version
- Cost of thoroughness
- Anthropic reported its multi-agent research uses about 15 times more tokens than a normal chat
- Known limits
- OpenAI says it can hallucinate facts, may struggle to tell rumours from authoritative information, and often fails to convey uncertainty
As of October 2026, deep research modes are part of ChatGPT, Gemini, Claude and Perplexity, and some can also search your connected files and work apps.
Questions people ask
What is ChatGPT deep research?
A mode in ChatGPT, launched in February 2025, where the AI spends several minutes browsing the web, reading sources and analysing them, then delivers a cited report. It is meant for questions that need many sources, not quick facts.
How is deep research different from normal AI search?
Normal search answers run one or a few searches and reply in seconds. Deep research plans, runs many rounds of searching and reading, follows leads, and writes a longer structured report, which takes minutes.
Can deep research reports be trusted?
They are a strong starting point, not a final answer. Reports can rest on weak sources, misread a page or present guesses confidently. Check the citations behind important claims, especially numbers, dates and anything you will act on.
Which deep research tool is best?
It depends on the question and changes with each update, so we don't rank them. Try the same question in two tools and compare their sources. Check whether a tool can search your own files, if that matters to you.