Gemini 3 is almost as good as Google says it is

Google set the bar high for Gemini 3. It’s promising a bunch of upgraded features in its shiny new AI model, from generating code that produces interactive 3D visualizations to “agentic” capabilities that complete tasks. But as we’ve seen in the past, what’s advertised doesn’t always match up to reality. So we put some of Google’s claims to the test and found that Gemini 3 delivers reasonably well — with caveats.

Google announced the Gemini 3 family of models earlier this week, with the flagship Gemini 3 Pro rolling out to users first. Gemini 3 Pro is supposed to come with big upgrades to reasoning, along with the ability to provide more concise and direct responses compared to Google’s previous models.

Some of the biggest promised improvements are to Canvas, the built-in workspace inside the Gemini app, where you can ask the AI chatbot to generate code, as well as preview the output. When building in Canvas, Google says Gemini 3 can interpret material from different kinds of sources at the same time, like text, images, and videos. The model can handle more complex prompts as well, allowing it to generate richer, more interactive user interfaces, models, and simulations, according to Google. The company says Gemini 3 is “exceptional” at zero-shot generation, too, which means it’s better at completing tasks that it hasn’t been trained on.

For my first test, I tried out one of the more complex requests that Google showed off in one of its demos: I asked Gemini 3 to create a 3D visualization of the difference in scale between a subatomic particle, an atom, a DNA strand, a beach ball, the Earth, the Sun, and the galaxy, as shown here.

Gemini 3 created an interactive visual similar to what Google demonstrated, allowing me to scroll through and compare the size of different elements, which appeared to correctly list each one from small to large, starting at the proton and maxing out at the cosmic web. (To be fair, I’d hope Gemini could figure out that a beach ball is much smaller than the Sun.) It included almost everything shown in the demo, but its image quality fell short in a couple of areas, as the 3D models of the strand of DNA and beach ball were quite dim compared to what Google showed. I saw much of the same when feeding Google’s other demos into Gemini. The model spit out the correct concept, but it was always a little shoddier, whether it had lower resolution or was just a little more disorganized.

Gemini 3’s output didn’t quite stack up to Google’s demo when I tried something a little simpler, either. I asked it to re-create a model of a voxel-art eagle sitting on a tree branch, and while my results were quite similar to the demo, I couldn’t help but notice that the eagle didn’t have any eyes, and the trees were trunkless. Branching out from Google’s example, a voxel-style panda came out alright, but standard 3D models of a penguin and turtle came out quite primitive, with little to no detail.

But Gemini 3 isn’t just built for prototyping and modeling; Google is testing a new “generative UI” feature for Pro subscribers that packages its responses inside a “visual” magazine-style interface, or in the form of a “dynamic” interactive webpage. I only got access to Gemini 3’s visual layout, which Google showed off as a way to envision your travel plans, like a three-day trip to Rome.

When I tried out the Rome trip prompt, Gemini 3 presented me with what looked like a personalized webpage featuring an itinerary, along with options to customize it further, such as whether I’d prefer a relaxed or fast-paced vacation or if I prioritize certain dining styles. Once you submit your preferences, Gemini 3 will redesign the layout to match your selections. I found that this feature can provide interactive guides on other topics, too, like how to build a computer or set up an aquarium.

Next up, I did a little experimenting with Gemini Agent, a feature Google is testing for Ultra subscribers inside the Gemini app. Like other agentic features, Gemini Agent aims to perform tasks on your behalf, such as adding reminders to your calendar and creating reservations.

One example shared by Google shows Gemini Agent organizing a Gmail inbox, so I asked the tool to do the same — and, well, it followed my orders. It found 99 unread emails from the last week and displayed them inside an interactive chart. Gemini provided options to set up reminders for what appeared to be the most important ones, such as RSVPs and a bill, while offering buttons to archive emails it identified as promotions. I asked Google Gemini to schedule a reminder to pay my bill, and the AI assistant put it inside Google Tasks with the correct due date. When I asked it to pay the bill, it navigated the billing interface and came close to asking me to enter my payment details, but (given the security concerns around agentic AI) I stopped short of letting it proceed.

While you could just organize your inbox manually, I found Gemini 3’s assistance somewhat helpful, as it dug up a few forgotten emails that I might’ve missed. You can also ask Gemini to find and unsubscribe from spammy email providers in bulk, which is nice.

Between Perplexity’s AI assistant, ChatGPT, and Gemini, Google’s AI chatbot (predictably) offers the richest integration with Gmail. Perplexity will pull up emails listed in your inbox, but you’ll need to tell it which ones to keep, archive, or delete, instead of just hitting a button like you can with Gemini. For some reason, ChatGPT refused to organize my inbox, claiming its integration with Gmail is in a “read-only” mode despite readily sending an email through the app on my behalf. But while Gemini is directly connected to Gmail, it was still far slower at sending emails in the app when compared to Perplexity.

Gemini almost managed to book a restaurant reservation without intervention, only to falsely tell me that there’s a “cost” associated with making the booking right before finalizing it. When I asked about the charge, Gemini 3 backpedaled and said it “likely referred to” the restaurant’s 16 percent service charge. It then proceeded to ask me to confirm my reservation three times and then told me there was a financial transaction involved again. Sigh. Again, I felt that I could complete these tasks far faster myself.

Despite the hiccups with completing tasks, Gemini 3 Pro’s interactive visualization features were impressive, and I could see how interactive models or visual layouts could be useful in some scenarios — though I can’t see myself using them on a daily basis, and Gemini’s text-based answers are usually informative enough for me. For now, I think I’ll just keep using Gemini like I always do: for questions I might not immediately find by browsing the web.