We’ve tested both Google’s and OpenAI’s Deep Research agents, so let’s see how they compare. OpenAI charges $20 for its Deep Research agent, while Google offers it for free. Gemini’s Deep Research agent is powered by the new Gemini 2.0 Flash Thinking model, while ChatGPT’s Deep Research agent uses a refined version of the o3 model. However, it’s interesting to note that Google offers its agent for free.
Task #1: Research on China’s Emergence in AI
To put the Deep study agent in ChatGPT and Gemini to the test, I requested them to perform extensive study on China’s rise to prominence in artificial intelligence. I asked both agents to look into prominent AI businesses in China, government policies, competition with US laboratories, and other topics. ChatGPT Deep Research agent completed the work in 10 minutes and cited 30 sources, yielding a 9,000-word report. In contrast, Gemini’s Deep Research agent spent 8 minutes researching and produced a much shorter 3,000-word report. However, it analyzed over 170 websites, which is impressive.
As for the research output, I thoroughly analyzed both papers and found that ChatGPT’s Deep Research agent utterly omitted to highlight China’s recent AI breakthroughs. It didn’t even discuss DeepSeek R1, Baidu’s new Ernie 4.5 model, and more. As it turns out, ChatGPT’s Deep Research agent relied largely on a Stanford article and a Wikipedia page which were written and last updated in 2017, and 2021, respectively. As a result, much of the information was outdated. It didn’t even mention the latest video generation models or robotics companies.
While being concise, Gemini’s Deep Research agent remarked, “Notably, DeepSeek, a startup established in 2023, has quickly risen to prominence with its R1 model.” In addition, Gemini added: “DeepSeek has also distinguished itself through its pioneering work in developing novel architectural advancements, such as Multi-Head Latent Attention (MLA).”
It also discussed video generation AI tools such as Kling AI, MiniMax, and AI-driven robotics companies like Unitree. My assessment is that Gemini’s advantage is Google’s search index, which is regularly updated with new web pages on every topic. OpenAI is likely to rely on Bing to curate webpages for research work, and as a result, it’s curating old information.
Task #2: Research on The Future of AI Chips over Nvidia GPUs
In the second challenge, I asked the Deep explore agent on ChatGPT and Gemini to explore the future of dedicated AI processors over Nvidia GPUs that are now employed for training and more. ChatGPT included 22 sources and generated a huge paper with tables and in-line citations. Gemini, like before, evaluated over 100 web sites but produced a brief report.
Both agents cited the rise of TPUs, ASICs, FPGAs, LPUs, and WSEs, as well as Nvidia’s CUDA moat. They talked about new silicon firms such as Groq, Cerebras, SambaNova, Graphcore, and others. ChatGPT’s Deep Research AI agent also brought up Huawei’s Ascend AI processor, which Gemini overlooked.
The agent is powered by a finely tuned version of OpenAI’s o3 model, with a training deadline of October 2023. As a result, the internal model does not reflect the most recent understanding. Despite leveraging the internet for current information, it frequently misses out on the most recent breakthroughs.
Task #3: Research on Obesity Management in 2025
In the third task, I directed the Deep Research agent on ChatGPT and Gemini to investigate obesity management in 2025. I explicitly asked both agents to include the most recent breakthroughs in 2025, encompassing all types of therapy. ChatGPT’s Deep Research quoted high-quality sources and medical journals to describe contemporary obesity management therapies.
It discussed new breakthroughs such as GLP-1 peptides, their side effects, FDA approval status for new medications, and even gene therapy. Overall, ChatGPT supported the present obesity management policies in the United States, United Kingdom, and Canada. On the other hand, Gemini’s Deep Research agent did an excellent job. It presented all medical treatments and forthcoming pharmacological trials in a table manner. What I found intriguing was that Gemini referenced novel medications such as the triple hormone receptor (Retatrutide), which is causing a stir in the medical community for significant weight loss.
Furthermore, Google’s Gemini dug deep into companies like Novo Nordisk and Eli Lilly, which are working with novel medications for weight loss, providing a comprehensive picture of the evolution. Overall, I believe Gemini’s report was well-structured and contained up-to-date material on obesity management.
Should You Use ChatGPT Deep Research or Gemini Deep Research?
During my testing, I liked Gemini’s Deep Research AI agent since it constantly provided current knowledge on a variety of themes. While ChatGPT’s Deep Research agent covers a wide range of topics, its poor understanding of the most recent developments prevents it from providing a comprehensive picture. Furthermore, Gemini’s Deep Research AI agent is free for all users, but OpenAI’s agent costs $20 to access. Gemini’s improved performance is due to the Gemini 2.0 Flash Thinking model. Compared to the prior Gemini 1.5 Pro model, the new reasoning model considers and plans the information required to accomplish the research activity.
However, one significant advantage of ChatGPT’s Deep Research agent is the ability to submit files, whereas Gemini does not enable file uploads when utilizing its agent. This is especially useful in STEM-related research projects, where you can submit PDFs of paywalled science periodicals. Because many of these publications are behind a paywall and unreachable via web search, you can manually contribute files to build a more comprehensive knowledge base. This enables the agent to absorb relevant insights. However, Gemini allows you to export the created report to Google Docs, which is useful for sharing.
Overall, I recommend using Gemini’s Deep Research AI agent before subscribing to the ChatGPT Plus package. Google has considerably enhanced the agent, and it now performs much faster.