What is Deep Research?
Deep Research is OpenAI’s new agent that can perform multi-step research on the internet for performing complex tasks like generating reports and competitor analysis.
- It has access to tools (Python) and the internet (browser tools).
- It is powered by the OpenAI o3 model, which is optimized for web browsing and data analysis. It uses reasoning to search, interpret, and analyze massive amounts of information.
- The model was trained using reinforcement learning (RL) to browse, reason for complex information, and learn to plan and execute multi-step trajectories to find the data it needs.
- It can also backtrack, adopt its plan, and react to real-time information as it needs.
- It supports user-uploaded files, generates plots (using Python), and embeds generated graphs and images from websites (not functional at the moment), including adding citations.
What problem does it solve?
Deep Research can perform complex multi-step research tasks in less time than humans would (minutes instead of hours).
Search + Analyze + Synthesize
- Report Generation
- Produce Insights
- Action Plan
It can do this on top of hundreds of online sources.
What are the use cases?
Professional Applications:
- Finance: Market and competitive analysis
- Scientific research and data analysis
- Policy and regulatory research
- Engineering documentation and analysis
Shopping & Consumer Research:
- Detailed product research (cars, appliances, furniture)
- Hyper-personalized recommendations
- In-depth product comparisons
Academic & Analysis:
- Literature review and comprehensive summaries
- Generate overviews with findings and discover new insights
- Identify research gaps → new research questions → novel scientific research
- Discover trends and find new recommended readings
- Analyzing quantitative outputs and generating interesting discussions
- Source verification and discovering new evidence
- Hypothesis testing?
Knowledge Work/Engineering:
- Answering complex queries requiring multiple steps
- Analyzing uploaded files and documents and augmenting with new research
- Creating comprehensive reports
- Developing technical documentation
- Conduct feasibility studies
- Synthesizing information from multiple sources
Examples:
- Recent Papers on o1 and DeepSeek-R1 (Summary & Analysis)
Deep Research excels at tasks that typically take humans many hours to complete, especially those requiring:
- Integration of multiple information sources
- Deep analysis of complex data
- Creation of well-documented reports
- Multi-step research processes (involving planning, finding, browsing, reasoning, analyzing, and synthesizing)
- Processing, understanding, and reasoning about large amounts of information
How to decide when to use Deep Research?
Use Deep Research if the task requires multi-faceted, domain-specific queries requiring extensive research for real-time information and careful reasoning/understanding.
For other tasks:
- Use o1-mini for tasks that benefit from reasoning (breaking down complex tasks into smaller parts in an autonomous way).
- Use GPT-4o for all other one-off simple tasks.
Usage Tips for OpenAI’s Deep Research
The more compute, the better results:
The more you allow the model to browse and think, the better the performance.
Prompt Engineering
- Clear and specific instructions: Provide a structured plan and be as detailed as possible.
- Clarify, don’t ignore: Answer clarifying questions to refine the request.
- Keywords help a lot: Use precise terminology (e.g., brand, technical term, product name) to reduce unnecessary processing.
- Use clear verbs: E.g., “compare,” “suggest,” “recommend,” and “report” to define expectations.
- Output Format: Specify report format, sections, tables, or layout.
- Upload files as context: PDFs and other documents provide important domain-specific knowledge.
- Check sources & verify information: Always double-check sources since the model can still make mistakes.
What to try?
Research:
- Market research/competitor analysis on AI tools
- Research around new products (reviews, price comparisons, etc.)
- Augment documents with additional details or critiques
- Extensive research for product feature recommendations
- User studies
- Legal case research (case laws, precedents, regulations)
- Fact-checking or background checks
Business Use Cases:
- Search and develop AI/agent use cases for a specific domain
- Track trends in a specific domain or topic
Learning Use Cases:
- Build a study plan and recommend a learning path
- Collection of coding tips and best practices for AI models
- Check for the latest features of developer tools and suggest exercises
Science:
- Latest research on health-related topics (sleep, symptoms, mental health, etc.)
- Write technical reports with the latest findings on a topic
Content Creation:
- Write a blog post on a combination of topics
- Suggest topics based on trends
- Generate slides for presentations
Personal:
- Develop a detailed biography of yourself or a public figure
- Develop/update a resume based on public information and projects
What’s missing/challenges?
- Struggles to synthesize technical information: Supporting documents help.
- Hallucinations: The model still makes mistakes and may struggle to distinguish authoritative information.
- Varying accuracy across domains.
- Challenges combining diverse information types.
- Lack of explicit control over online source selection.
- Citation mistakes and formatting errors.
- Limited export options: Cannot yet export directly to Excel, Notion, Docs, etc.
- Not great with time/date-related queries.
- Paywalled/subscription-based sources unsupported (future integrations possible).
- Chart embedding not functional yet.
- Cannot take actions yet: Deep Research primarily reads and analyzes but does not interact with external systems beyond searching.
Future Possibilities
- Site search capabilities (e.g., advanced search on arXiv)
- More tools and access to knowledge bases
- Enhanced personalization via custom instructions
- Better citation handling and formatting improvements