How do you get the most from Gemini 3 in the Gemini app?

Gemini 3 Pro reached the Gemini app on November 18, together with a redesigned interface. It is easy to keep typing the same short questions as before, but that leaves most of the upgrade unused. The model's strengths are reasoning through a messy problem, reading images and documents, and building small interactive things on request. To get those, you have to choose the right mode and hand it real material.

This guide covers the settings that matter, three worked tasks you can copy, and the checks that keep the results trustworthy.

Step 1: Pick the right mode

In the model selector at the top of a chat, choose Thinking. That is how Google exposes Gemini 3 Pro in the app; it is rolling out to users worldwide, while Google AI Plus, Pro and Ultra subscribers get higher usage limits. "Thinking" means the model works through the problem before it writes, so answers take a little longer. For a quick definition or a one-line rewrite, a faster option in the same menu is fine. For anything with several steps, constraints or attachments, use Thinking.

Two other places in the redesigned app are worth knowing. My Stuff collects the images, videos and reports you have made, so you can find them again. Canvas is the side workspace where Gemini writes documents and code that you can edit and run.

Step 2: Give it material, not just questions

Google's model card describes Gemini 3 Pro as natively multimodal, meaning it was trained to take in text, images, audio and video together, and lists a context window of up to a million tokens (a token is a word fragment, roughly three-quarters of a word). In practice, the more of your actual situation you upload, the less it has to guess.

Here are three tasks that show the difference. The prompts are ready to use; the outcomes described are illustrative.

A photo to a table. Photograph a page of handwritten notes, a whiteboard or a stack of receipts, attach it, and write:

Read this photo carefully. Turn it into a table with one row per item:
date, description, amount. Mark any word or number you cannot read
with [?] instead of guessing. Then give the total of the amount column.

The [?] instruction matters. It turns a silent misreading into a visible gap you can fix.

A small tool in Canvas. Google calls Gemini 3 its strongest model yet for "vibe coding", which means building software by describing what you want in plain language. Open Canvas and try:

Build a simple web page that helps a family split grocery costs.
Inputs: each person's name and what they paid. Output: who owes whom,
with the fewest payments possible. Large buttons, works on a phone.
Explain in three sentences how the calculation works.

Run it in Canvas, enter numbers you can check in your head, and ask for one change at a time.

A plan with a visual layout. Google is testing two generative interfaces, layouts the model designs for each answer. Visual layout builds a magazine-style page with photos and modules; dynamic view writes a small interactive page on the spot. You may see only one of them while Google compares the two. A trip or learning plan is a good test:

Plan three days in Quebec City in late December for two adults who
like history and food but not long walks in the cold. Budget-conscious.
Include indoor options for each afternoon.

Step 3: Check what it cannot know

The model card gives Gemini 3 Pro a knowledge cutoff of January 2025: what it learned in training stops there. For prices, opening hours, schedules or recent news, ask it to search and to show where each fact came from, then open those links. Google also lists hallucinations, meaning confident but false statements, among the model's known limits. A simple habit helps: end important requests with "List anything in your answer you are unsure of."

For agent-style work, Google is starting with Gemini Agent, an experimental mode for multi-step tasks such as sorting an inbox or preparing a car rental from details in your email. For now it is limited to U.S. subscribers of Google AI Ultra, on the web, and it is designed to ask for confirmation before buying anything or sending messages.

Which tasks suit it best?

Tasks that combine reading with reasoning suit it best. Examples include turning a photographed document into structured data, explaining a chart or a homework problem from an image, summarizing a long report against your own questions, and comparing several options against a list of constraints. Google's own numbers point the same way: it reports strong scores on tests of image and video understanding and on hard reasoning questions, and it highlights the model's coding ability for building interactive pages.

Small builds are another good fit. Calculators, checklists, quizzes and simple visual explainers made in Canvas let you see quickly whether the result works, which is the best kind of task for any AI model.

It suits fresh facts less well unless you ask it to search and you check the links. It also suits less well any job where a mistake is costly and hard to spot, like legal wording or medical dosages, unless a qualified person reviews the result. Use Thinking for depth, attach your real material, and keep the final judgment for yourself.