 |
The Google Gemini logo. Photo: Bloomberg . |
Google launched the Gemini 3 model in late November. The company introduced a series of notable upgrades such as improved coding capabilities, visual image creation, and agents that perform certain tasks on behalf of the user.
When introducing Gemini 3, Google emphasized that it was the company's "smartest model yet." Real-world experience shows that the model processes faster and has better reasoning capabilities, although there are still some areas for improvement.
Create interactive interfaces quickly.
One of the major upgrades in Gemini 3 revolves around Canvas. This is an interactive workspace within Gemini that allows you to ask chatbots to write code, visualize results, and edit content.
When Gemini 3 is asked to create a 3D model of the size differences between a soccer ball, the Earth, the Sun, and the Milky Way, the tool activates a coding interface, processes and creates an interactive environment, allowing users to select and compare the sizes of objects.
 |
The ability to create interactive spaces using the Canvas feature in Gemini 3. |
Compared to Google's demo, the results are slightly inferior in image quality. However, the information about the objects is clear and easy to understand. Processing time is approximately 2 minutes. In the Canvas interface, users can directly edit the code, copy, or share the content.
With another command, Gemini 3 was asked to create a 3D cube-style (voxel) model of an eagle perched on a tree branch. The tool produced the result in about 45 seconds, and users could zoom in and rotate the view using the mouse. However, some details in the image had graphical errors and were not displayed stably.
The eagle perched on a tree branch was also present, but the requirement was more difficult (a 3D model). Initially, the code generated by Gemini 3 failed and wouldn't run. After requesting modifications, the image appeared and became interactive.
 |
Gemini 3's ability to create interactive environments. |
Despite some remaining issues, the shift from a question-and-answer model to a collaborative, editing approach is an interesting direction, making Gemini 3 useful for complex, long-term tasks.
Of course, Gemini 3 isn't just for creating graphics or visual models. Google is testing a "Generative UI" feature for AI Pro users. With certain commands, the model can create interactive magazine-style interfaces or visual websites.
Google says this feature is being rolled out in stages. According to The Verge 's experience, users need to select the Visual Layout feature, then enter a command such as "Plan my 3-day trip to Rome."
After processing, Gemini 3 generates results similar to a travel website, complete with itineraries, detailed images, and some questions about travel preferences and eating habits. The model then adjusts the answers accordingly. This feature is also quite useful for topics such as how to assemble a computer or build an aquarium.
More effective reasoning and analysis.
According to Google, Gemini 3 has better reasoning capabilities than its predecessor. The company claims that Gemini 3 Pro's reasoning ability reaches a doctoral level, scoring 37.5% on the Humanity's Last Exam and 91.9% on the GPQA Diamond exam.
Of course, users can only truly appreciate the difference when interacting directly with the model. When asked to solve a problem from the GPQA Diamond reference question set (focused on doctoral level), Gemini 3 Pro took approximately 30 seconds to provide a detailed solution, including step-by-step instructions and illustrative images.
 |
Gemini 3 can answer complex questions, demonstrating thinking abilities equivalent to a PhD level. |
According to the assessment, the solution is quite detailed and follows the correct procedure. For comparison, GPT-5.1 also took about 30 seconds to answer correctly, but the solution was less detailed. Similarly, Claude 4.5 Sonnet answered correctly and faster than Gemini 3, but was less detailed.
Google says Gemini 3 can extract documents simultaneously from multiple sources, such as text, images, and videos . Tasks that have never been trained on before are also handled better on Gemini 3.
 |
Gemini 3's video analysis capabilities. |
For example, when asked to analyze the main topic of a 57-minute video, Gemini 3 took only about a minute to provide results including the requested arguments and timestamps to the corresponding segment of the video. Similarly, the model correctly identified a tennis video, analyzing the matches and the playing style of each athlete.
Users can ask Gemini 3 to simultaneously watch videos and refer to external sources, even asking questions for deeper analysis. Overall, Gemini 3 is quite responsive in analytical and document-referencing tasks, partly thanks to its connection with the Google Search database.
The agent is helpful, but there are still issues.
The Agent feature is also a notable upgrade on Gemini 3. For users of the AI Ultra plan (currently only available in the US market), this mode can perform complex tasks on behalf of the user, such as categorizing and organizing emails, searching for information, and booking travel tickets.
When launching Gemini 3, Google used the example of the ability to organize the inbox in Gmail. According to The Verge 's experience, this feature works exactly as requested: creating a complete list of unread emails for a week.
Similar to the Canvas feature, email lists and suggested actions are displayed in a separate area. Gemini offers several options such as setting reminders for important emails, archiving promotional emails, etc.
 |
Gemini 3's agent feature helps manage Gmail inboxes. Image: The Verge . |
When prompted to create a bill payment reminder, Gemini quickly inputs the relevant information into Google Tasks. For the payment task, the agent automatically navigates and stops when it needs the user to enter payment information.
Overall, Gemini 3's agent is quite useful in detecting emails that users tend to forget. The tool can also find and unsubscribe from emails from advertising addresses, which are no longer useful.
In fact, the agents of some chatbots like Perplexity and ChatGPT already support integration with Gmail. However, Gemini offers the most diverse integration capabilities. For example, Perplexity users have to specifically enter the emails they want to keep or delete, while these actions can be performed via buttons on Gemini.
 |
The interactive interface creation feature of Gemini 3. Photo: The Verge . |
The email sorting feature on ChatGPT is also unstable. However, Gemini's email sending speed is slower than Perplexity. Some tasks, such as making restaurant reservations, sometimes encounter problems.
Although it always requires user confirmation before performing sensitive actions like payments or sending emails, delegating certain tasks to AI still makes users hesitant. The agent mode on Gemini 3 is quite useful, but users still need to monitor the output.
Source: https://znews.vn/gemini-3-tot-den-dau-post1607493.html
Comment (0)