What Is OpenAI's "Deep Research" and Why Does it Matter?
Summary
What is "Deep Research?" OpenAI has unveiled a new AI agent called "deep research" for ChatGPT, designed to conduct complex, in-depth research, a feature aimed at professionals in fields such as finance, science, policy, & engineering who conduct thorough research as part of their jobs. The tool is multi-modal, a.k.a., able to analyze multiple data formats, & includes both web browsing & integrated python tools for downloading & analysis to be integrated into its operation. Some examples of what deep research is most useful for includes queries such as conducting a competitive analysis in a specific industry, analyzing the historical trends of animal populations, or finding & analyzing shows based on vague plot details. When a user asks deep research a question the tool will browse the web & put together a formatted report, usually within 5 to 30 minutes.
We see key features of deep research include: 1) Autonomous operation, or the ability to browse the web itself to research multiple topics requested, 2) the Processing & Analysis of information from text, images, or PDFs, into readable, academically oriented reports, 3) Time Efficiency, & 4) Documentation of Sources through citation. This features capabilities are harness using the upcoming o3 AI model which OpenAI has been working on, which the 03 mini feature all users can use today. Differing from o3 mini as a general prompt & response service, we see deep research has a focus on its ability to browse the web & analyze historical data sets. In fact, the tool was specially trained through task-specific, end-to-end reinforcement learning on real-world tasks across various domains.
Why Now? OpenAI's release of deep research on the tail of last week's news that DeepSeek's R1 model only cost $5.8M to develop (though after further research the investment community speculates this only includes on final testing model, & not the hours spent researching existing models, the programming, nor the power & equipment used to download/analyze the utilized information set for training). Subsequently, investors speculate the announcement is a response to DeepSeek's R1 release, noting the feature, bring run on o3's developing cognitive model, attained an accuracy score of 26.6% on the Humanity's Last Exam benchmark, significantly outperforming other AI models such as DeepSeek R1's 9.4% score. We remind investors that OpenAI had been building anticipation for the release of its updated o3 AI model since late 2024, & consumer do not yet have full access to it, but all free users are able to use OpenAI's o3 mini model today.
Will Deep Research Be Used Differently From Prior AI Models? We find deep research interesting in that it is taking one step forward into the actions research analysts take when completing research. A consensus opinion among analysts that have tried to use AI models in their job note how the tool can provide useful summaries on information from reference data, but the analytical conclusions from these public models & the brief responses shared are lacking in depth & reference (a key reason many analysts find the citations from Anthropic's Perplexity web tool more useful than the summaries from ChatGPT). Further, many AI tools are woefully inadequate at responding with useful or unique numerical analysis in any form, instead simply regurgitating numbers from sources without any reference or reasoning to filter the response properly. While we've yet to get our hands dirty using Open AI's deep research feature, the idea of using a tool designed specifically for where other cognitive models has failed is extremely intriguing.
What Are Investment Implications? Investors should note that by bringing on its incremental tools and features, OpenAI is continuing to make the argument AI will continue to grow utilizing the cloud & large amounts of hardware to continue developing. Further, by taking a step forward & into the workflow process, AI is further ingratiating itself as a tool to be utilized by enterprises and individuals alike & will likely continue to expand its energy & hardware footprint as it participates more in our daily lives. As previously highlighted in our overview of the 1990s fiber build out to AI analogy bear case misses the point (From Fiber to Silicon: How AI's Appetite Differs From the 90s Telecom Snack) we continue to believe cloud & hardware names on our radar will continue to benefit from the move of the workload from physical to digital, including both digital infrastructure & cloud computing providers.
Click here to access the full report and Jesse’s insights on the D. Boral Capital Research Portal: https://sso.bluematrix.com/idp/authn?conversation=e1s1&OCT=A32B-LBBF-ERCR-62QD
—
Analysts on this report
Jesse Sobelson, CFA