OpenAI's Deep Research: A Guide With Practical Examples

OpenAI has just released Deep Research, an AI agent powered by a version of the upcoming o3 model. It’s designed to browse the web, analyze multiple sources, and synthesize large amounts of information.

You might be wondering: Doesn’t ChatGPT already do this?

Unlike a regular ChatGPT session that generates quick responses, Deep Research can conduct multi-step investigations, reference multiple sources, and produce structured reports.

For instance, if you’ve ever researched the best car to buy—comparing reviews, weighing costs, etc.—you know that finding reliable information takes time and a lot of Internet browsing. Deep Research is built for exactly this kind of work.

I’ve tested Deep Research, and I’ve been both amazed and disappointed. It shows great potential, but it also produces incorrect facts and inferences. In this blog, I’ll be your human agent and summarize everything you need to know about Deep Research. I’ll walk you through practical examples, share prompting tips, and show you where Deep Research shines—and where you need to be extra careful.

What Is OpenAI’s Deep Research?

OpenAI’s Deep Research is an AI-powered agent designed to conduct in-depth, multi-step research on the Internet. Unlike standard ChatGPT browsing capabilities, which provide quick responses, Deep Research autonomously finds, analyzes, and synthesizes information from hundreds of online sources.

Deep Research is designed for anyone who needs comprehensive and reliable research, including:

• Professionals in finance, science, policy, and engineering who need well-cited, structured reports
• Business strategists conducting competitive analysis or trend forecasting
• Researchers and students gathering information from multiple sources
• Shoppers and consumers making high-stakes purchasing decisions (e.g., cars, appliances, real estate)
• Writers, journalists, and analysts who require fact-checked, multi-source insights

Essentially, if a task involves browsing multiple sources, cross-referencing data, and synthesizing information into a useful format, Deep Research is the tool for the job.

How Does Deep Research Work?

Powered by a version of the upcoming o3 model, Deep Research builds on OpenAI’s advancements in reasoning models but is specifically optimized for web browsing and real-world data analysis.

To achieve this, OpenAI trained Deep Research using reinforcement learning on real-world browsing and reasoning tasks. This allows the model to follow an iterative, step-by-step research process, improving its ability to synthesize complex topics into structured reports.

Deep Research Benchmarks

Humanity’s Last Exam

Humanity’s Last Exam is a newly released benchmark designed to test AI on expert-level multiple-choice and short-answer questions across over 100 subjects, from linguistics and rocket science to ecology and mathematics. This evaluation measures an AI’s ability to reason across disciplines and seek out specialized knowledge when needed—a critical skill for research-oriented models.

Deep Research achieved a record 26.6% accuracy, far surpassing previous models, including OpenAI’s own o1 (9.1%), DeepSeek-R1 (9.4%), and Claude 3.5 Sonnet (4.3%). Notably, the largest improvements over OpenAI’s o1 were seen in chemistry, humanities and social sciences, and mathematics, where Deep Research demonstrated its ability to break down complex questions and retrieve authoritative information.

* Models tested on the text-only subset of the exam because they are not multimodal. Source: OpenAI

GAIA

GAIA (General AI Agent benchmark) evaluates how well AI systems handle real-world questions, requiring a combination of reasoning, web browsing, multimodal fluency, and tool-use proficiency.

Deep Research set a new state-of-the-art (SOTA) record, leading the external GAIA leaderboard with strong performance across all difficulty levels. The model showed particularly high accuracy in Level 3 tasks, which require complex, multi-step research and synthesis.

Source: OpenAI

Deep Research’s high pass@1 score shows that even its first attempt at answering a GAIA question is more accurate than previous models. The cons@64 score (which measures performance with multiple response attempts) further highlights its ability to self-correct and refine its answers based on new information.

Internal evaluations

OpenAI also performed internal evaluations, in which Deep Research was rated by domain experts on expert-level tasks. I found the internal evaluations quite interesting!

The graph below shows that the model’s pass rate increases as it makes more tool calls. This highlights the importance of letting it browse and analyze information iteratively—giving it more time to think leads to better results.

Source: OpenAI

Let’s take a look at another graph—see below. Deep Research performs best on tasks with lower estimated economic value, with accuracy dropping as the task’s potential financial impact increases. This suggests that more economically significant tasks tend to be more complex or rely on proprietary knowledge that isn’t widely accessible online.

Source: OpenAI

The graph below contrasts pass rates with the estimated hours a human would take to complete each task. The model performs best on tasks that would take 1-3 hours for a person, but performance doesn’t decline consistently with time—indicating that what AI finds difficult doesn’t always align with what humans find time-consuming.

Source: OpenAI

How to Use Deep Research: Practical Examples

At the time of publishing this article, Deep Research is only available to Pro users, with a limit of 100 queries per month, but OpenAI plans to expand access to Plus, Team, and Enterprise users soon.

In my opinion, Deep Research is still in its early phase. While it shows great promise, the first example below highlights many of its problems. However, the second example demonstrates its immense potential.

Example 1: AI ecosystems

I constantly struggle to get a complete overview of the AI ecosystems of different companies. Take Google, for example—they have Gemini 2.0 Flash, Imagen 3, Veo 2, Project Mariner, Project Astra… what else am I missing? To finally get a clear overview, I prompted OpenAI’s Deep Research with this request.

Note that the model didn’t begin the research directly but instead asked for clarifications. In all my tests, no matter how specific my first prompt was, the model has always attempted to narrow down its research. This is useful, in my opinion, because I often think my prompt is clear and specific, but it usually benefits from some refinement.

I answered the model’s questions, and then the research began. A panel opened on the right side of the browser, showing the agent’s activity and sources in real time: