OpenAI has introduced deep research, a new AI-powered research assistant inside ChatGPT that allows users to conduct structured, multi-step investigations instead of receiving instant AI-generated responses. Designed for professionals in fields such as finance, science, and policy, the feature enables more complex research workflows by analyzing information across multiple sources and compiling findings into detailed reports.
Unlike standard ChatGPT interactions, deep research takes between five and thirty minutes to complete each query, depending on complexity. OpenAI has positioned the tool as a way to assist users with tasks that require methodical data gathering, such as policy analysis, technical market reports, and major purchase comparisons.
The feature is currently available exclusively to ChatGPT Pro users, who can make up to 100 queries per month. OpenAI plans to expand access to Plus, Team, and Enterprise users in the near future, but no timeline has been provided for availability in the UK, Switzerland, or the European Economic Area.
How OpenAI’s Deep Research Works
ChatGPT Pro users can activate deep research by selecting the option in the ChatGPT message composer and submitting a query, optionally attaching files such as spreadsheets or PDFs for added context.
Once initiated, the AI independently scans online sources, interprets key data, and compiles a structured response. A sidebar provides a step-by-step breakdown of its research process, showing how the AI gathered and analyzed information.
Initially, deep research outputs will be text-only, but OpenAI has confirmed that future updates will introduce embedded images, charts, and other visual elements. The company also plans to extend deep research’s capabilities by allowing it to pull from subscription-based and proprietary databases, making it more useful for business and enterprise applications.
Deep research is part of a larger vision OpenAI has for AI-driven productivity. The company recently introduced Operator, an AI tool designed to perform actions on behalf of users, such as booking appointments, making purchases, or automating online workflows.
Built on OpenAI’s o3 Model, Not GPT-4o
Unlike ChatGPT’s standard interactions, which rely on OpenAI’s GPT-4o model, deep research is powered by a specialized version of OpenAI’s o3 model. This model has been optimized for research-driven tasks, focusing on long-form analysis, browsing capabilities, and multi-step reasoning.
OpenAI states that “deep research was trained using reinforcement learning on real-world tasks requiring browser and Python tool use”, allowing it to adapt dynamically to complex research requests. Reinforcement learning enables AI models to refine their approach through iterative feedback, ensuring improved accuracy and efficiency over time.
However, OpenAI acknowledges that deep research is not flawless. The company warns that the AI may “struggle with distinguishing authoritative sources from rumors” and that it “often fails to convey uncertainty accurately” (OpenAI). Additionally, citation formatting errors and inconsistencies may appear in its research reports, an issue that OpenAI says it is actively working to improve.
Google’s Competing AI Research Assistant
OpenAI’s deep research isn’t launching in a vacuum. Google recently announced its own AI tool called Deep Research, which is expected to be integrated into the Gemini AI ecosystem. Unlike OpenAI’s tool, Google’s version remains in a research phase and has not yet been made available to the public. However, Gemini Advanced users can test it already together with Google’s Gemini 2.0 Flash Experimental model.
Google’s Deep Research is part of its broader Project Mariner, an initiative aimed at developing AI agents that can autonomously browse the web, synthesize information, and create structured research reports. The parallel development of these tools highlights a growing trend: AI is moving beyond simple chatbot interactions and toward fully autonomous research assistants.
Benchmark Performance Against Other AI Models
To measure its effectiveness, OpenAI tested deep research against Humanity’s Last Exam, an AI benchmark designed to assess performance on 3,000 expert-level questions. Deep research scored 26.6%, far exceeding competing models:
• Google’s Gemini Thinking: 6.2%
• Grok-2: 3.8%
• Claude 3.5 Sonnet: 4.3%
• OpenAI’s own GPT-4o: 3.3%
Additionally, deep research set a new high score on the GAIA benchmark, which evaluates AI reasoning across real-world multimodal tasks. While these scores position deep research as one of the most advanced AI research tools available, its real-world effectiveness remains to be fully tested.
Limitations and Early Challenges
Despite its potential, deep research comes with several limitations that OpenAI has openly acknowledged. One of the primary concerns is accuracy. While the tool provides citations for its sources, OpenAI has warned that it “may struggle with distinguishing authoritative sources from rumors” and that it “often fails to convey uncertainty accurately” (OpenAI). This means users must still verify information manually to ensure credibility.
Another limitation is that deep research is currently restricted to open web browsing and user-uploaded files. It does not yet have access to proprietary knowledge bases or subscription-only content, such as academic journals behind paywalls. For businesses that rely on internal data or industry-specific reports, this may reduce the tool’s usefulness compared to human researchers with direct access to such resources.
Formatting inconsistencies are another issue. OpenAI has admitted that deep research may produce citation errors, including misattributions or broken reference structures. This is an important factor for professionals relying on precise documentation, and OpenAI has stated that improvements in citation handling are in development.
Security Risks in AI-Powered Research
Beyond accuracy concerns, the launch of deep research also raises broader questions about the security of AI-driven research tools. AI-generated misinformation has been a recurring issue in AI-assisted search models, and OpenAI has acknowledged that deep research is not immune to hallucinations, where the model generates incorrect or misleading statements.
Recent security incidents involving other AI models suggest that these risks are not theoretical. For example, China-based AI company DeepSeek faces scrutiny after an independent AI audit revealed that its DeepSeek R1 model failed 83% of accuracy tests. The report cited issues with misinformation and potential political bias in AI-generated research.
Cybersecurity researchers have also exposed vulnerabilities in AI research tools that interact with online sources. A report from Palo Alto Networks’ Unit 42 revealed multiple ways to jailbreak DeepSeek’s AI, manipulating its responses to provide misleading or harmful information. While there are no documented exploits affecting OpenAI’s deep research yet, the increasing sophistication of AI jailbreak techniques presents a potential security concern.
These challenges highlight the broader risks of using AI for research. While deep research aims to provide more structured and verifiable results than past AI search tools, it is not a flawless substitute for human researchers who can critically evaluate sources and detect unreliable information.
What Comes Next for OpenAI’s Deep Research?
OpenAI has described deep research as the first step toward more capable AI assistants that can conduct autonomous investigations. However, the company has also emphasized that the feature is highly resource-intensive, which is why access is currently restricted.
At launch, deep research is only available to ChatGPT Pro users, with a limit of 100 queries per month. OpenAI has stated that Plus and Team users will be next in line for access, followed by Enterprise customers.
The company has also acknowledged that deep research will need further refinement. OpenAI is actively working on improving the accuracy of citations and formatting in research reports. Additionally, the AI will eventually support access to proprietary and subscription-based data sources, which would make it significantly more useful for enterprise users who rely on high-quality industry reports rather than open web content.
A key milestone for deep research is its upcoming availability on mobile and desktop apps. OpenAI has confirmed that the feature will roll out beyond the web version of ChatGPT in the coming weeks, providing broader accessibility to users who need research assistance on different platforms.
AI Research as a New Competitive Battleground
OpenAI and Google’s simultaneous push into AI-powered research tools suggests that automated knowledge synthesis is becoming a new frontier in AI development. With deep research, OpenAI is attempting to move beyond traditional chatbot interactions into AI-assisted research workflows, where users receive structured reports instead of conversational responses.
The competition between OpenAI, Google, and Microsoft suggests that AI research assistants could become an essential part of future AI offerings. Whether these tools will ultimately replace traditional research methods or serve as complementary aids remains to be seen.
The integration of deep research and the Operator AI Agent suggests that OpenAI is moving toward an AI ecosystem where ChatGPT is not just a conversational assistant, but a system capable of independently conducting research and executing tasks. This aligns with OpenAI’s long-term vision of creating AI agents that can function as assistants in real-world applications.
The competition in this space is intensifying. Google’s Gemini AI models are expected to incorporate similar AI agent capabilities, and Microsoft is also working on enhancing AI research functions within Azure. The question now is not whether AI-powered research assistants will become widely adopted, but rather how they will integrate into professional and consumer workflows.
Can AI Research Replace Human-Led Investigations?
The rise of AI-powered research tools raises an important question: can AI fully replace human-led investigations? While deep research provides structured, citation-backed reports, it does not yet possess the ability to critically evaluate sources in the way that human researchers can. AI models still struggle with bias, misinformation, and understanding context beyond what is explicitly written in source materials.
Even with OpenAI’s efforts to improve accuracy, the AI is ultimately limited by the data it can access. Academic and corporate researchers who rely on proprietary datasets and specialized knowledge may still find human-led investigations more reliable. Additionally, deep research lacks the ability to conduct interviews, interpret qualitative insights, or assess real-world implications beyond what is available in its sources.
That said, AI research assistants like deep research could serve as valuable tools for professionals who need to streamline the initial stages of information gathering. Rather than replacing human researchers, these tools may function as productivity enhancers that automate time-consuming tasks such as literature reviews, data comparisons, and report generation.
Meanwhile, Microsoft’s increasing investment in AI-powered enterprise research tools suggests that businesses will become a key market for these capabilities. OpenAI’s eventual expansion of deep research into corporate environments could make it a central part of AI-assisted knowledge work.
While deep research represents a step forward for AI-driven research automation, its current version still has limitations. The tool’s reliance on open web content, potential for misinformation, and query restrictions mean that it remains an evolving product rather than a fully mature solution.
Table: AI Model Benchmarks – LLM Leaderboard
[table “18” not found /]Last Updated on March 3, 2025 11:33 am CET
