Deep Research

Deep research is an agentic AI capability in which a model independently plans and executes a sequence of searches, reads and analyses hundreds of sources over tens of minutes, and returns a comprehensive, cited report — a task previously requiring many hours of a human analyst's time.^[1]

History

Google launched Gemini Deep Research in November 2024, presenting a research plan for the user to approve before the agent ran. OpenAI followed on 2 February 2025 with deep research in ChatGPT, powered by a browsing-optimised version of its o3 reasoning model, initially for Pro subscribers and expanded to Plus users on 25 February 2025.^[1][2]

Perplexity released its own Deep Research about two weeks later, and open-source reimplementations — most visibly Hugging Face's open deep research built on the smolagents library — arrived within the same month. Earlier precursors included the Allen Institute for AI's ScholarQA (January 2025) and academic systems such as STORM that generate cited long-form synthesis.^[2]

OpenAI extended the feature in stages: a lightweight o4-mini variant in April 2025 raised usage limits (25 queries per month for Plus users, 250 for Pro), and a February 2026 update added connectors to MCP servers and apps, restriction of searches to trusted sites, and real-time progress that users can interrupt and redirect.^[1]

Key Concepts

Deep research differs from a normal chatbot answer in three ways:

• Multi-step, iterative search. The model reformulates its query, follows citations, and pivots when a source turns out to be irrelevant — a loop of search, read, reason, re-search rather than a single retrieval pass.
• Reasoning under a budget. The agent spends inference time deliberating over what it has read, which is why deep research is backed by reasoning models rather than base chat models. OpenAI reported 26.6% accuracy on Humanity's Last Exam for deep research against 9.1% for o1 at launch.^[1]
• Grounded synthesis with citations. The final report maps claims to the sources consulted, making the output auditable in a way free-form generation is not.

The cost is latency and tokens: runs take minutes and consume far more compute than a standard query, which is why vendors meter deep research separately from ordinary chat usage.^[1]

Applications

Deep research is used for market and competitor analysis, literature reviews, vendor evaluation, due diligence, technical scouting and policy monitoring. Media and research organisations use it to prepare background briefings; students and analysts use it to map a field before reading primary sources. It complements AI search engines: a search engine answers a question in seconds, while deep research produces a structured document on a topic.^[3]

Its failure modes are familiar from other LLM systems — hallucinated citations, over-trust of low-quality web pages, and confident synthesis of contradictory sources — so outputs are generally treated as a first draft requiring human verification rather than a finished deliverable.^[2]

>See Also

• AI Search Engines
• Reasoning Models
• Agentic AI
• Perplexity AI
• Humanity's Last Exam