OpenAI has unveiled ChatGPT Deep Research, a new tool that promises to act like an AI analyst — scouring the internet, verifying sources, and coming up with structured answers. But is this really the future of research, or just another AI trick with the occasional hallucination?
OpenAI is apparently accelerating the development of a new generation agents – AI tools that not only answer questions, but independently research, analyze data, and present resultsThe latest addition to the family of these tools is Deep Research, a feature that is currently exclusive to ChatGPT Pro subscribers.
Rather than just giving the user a quick answer, Deep Research plans a multi-step research process. It shows key steps in its analysis in a sidebar, cites sources, and corrects its own errors when necessary. The goal is to provide more reliable and supported answers, similar to what a human researcher would do.
How does it work? (And why it takes up to 30 minutes!)
Deep Research allows users to upload text-based questions as well as PDFs, images, or Excel spreadsheets, which provide the AI with additional context for research. Then the system performs several sequential steps of data search and analysis, which can last from 5 to 30 minutes, depending on the complexity of the topic.
Key features of Deep Research include:
- Multi-level information search – AI does not provide an answer immediately, but gradually collects and verifies data from various sources.
- Automatic correction of incorrect paths – if the system is caught in incorrect reasoning, it can correct its methodology and recheck the data.
- Transparency of the research process – users can see in the sidebar what sources the AI used and how it reached its conclusions.
This all sounds promising, but OpenAI warns that the system still not perfectDeep Research can sometimes hallucinates, incorrectly assesses the credibility of sources, or provides answers with a low level of confidence. This means that users still need to critically evaluate and verify data.
How does Deep Research compare to the competition?
Other technology companies are also developing similar concepts. Google to launch Project Mariner in December 2024, an AI research prototype, but it is not yet available to the public. While Google remains in the experimental phase, OpenAI has already offered Deep Research to paying users.
A comparison of key features shows some interesting differences:
| Function | OpenAI Deep Research | Google Project Mariner |
|---|---|---|
| Accessibility | Available to Pro users | Still in testing phase |
| Speed | 5–30 min to respond | No data |
| Sources | Shows quotes and methods | Unknown |
| Data integration | Supports PDFs, images, tables | It is not known |
| Accuracy | 26.6 % on benchmark test | It is not known |
OpenAI seems to have an advantage because it offers working product, but the question remains how reliable and useful such AI research is in practice.
How much does it cost and who can use it?
Currently, Deep Research is available exclusively to subscribers ChatGPT Prowho pay 20 $ per month, but with a significant limitation 100 queries per monthOpenAI also plans to roll out access for Plus, Team, and Enterprise subscribers, promising better performance and speed in the future.
The main limitation remains high computational complexity, as Deep Research consumes significantly more computing resources than the classic version of ChatGPT. Therefore, OpenAI currently offers only a limited number of queries and promises optimizations for lower costs in the future.
Is Deep Research really replacing human analysts?
Deep Research is undoubtedly an important step towards AI researchers, but in practice it is still being used more as an aid to analysts, not as their replacement.
Its key advantages are:
- Structured and multi-stage research process, which provides better answers than a classic generative AI tool.
- Citing sources, which increases transparency and allows for verification of information.
- Adapting based on new data, as AI can correct its analysis if it encounters conflicting information.
However, it also has obvious limitations:
- AI can misinterpret data or hallucinate, which means that the answers are not always reliable.
- Speed is not comparable to classic ChatGPT, as the surveys take between 5 and 30 minutes.
- Human supervision is still needed, as AI cannot separate all sources as well as an experienced analyst.
Although Deep Research is a promising step forward, it cannot yet completely replace human research – at least not yet.
Would you pay 200 $ per month for an AI assistant to research for you?
