# What is Gemini 3 Pro, and what is new in Google's latest model?

Date: 2025-11-19 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gemini-3-pro/

**Answer:** Gemini 3 Pro is the first model of Google's Gemini 3 generation, released in preview on November 18, 2025, and Google describes it as its most intelligent model for reasoning, understanding images and video, and coding. It arrived the same day in the Gemini app, in AI Mode in Search, and in Google's developer tools, alongside a new agent-first coding platform called Google Antigravity.

Yesterday, Tuesday, November 18, 2025, Google released **Gemini 3 Pro**, the first model in its third Gemini generation. In a joint post, Sundar Pichai, Demis Hassabis and Koray Kavukcuoglu call it Google's most intelligent model, built to understand not only what you ask but the intent behind it, so you need less back-and-forth to get a useful answer.

What stands out is the scale of the launch. For the first time, Google put a new Gemini model into **Search on day one**, through AI Mode, its chat-style search experience. The same day, Gemini 3 Pro reached the Gemini app, developers and businesses.

## What Google released

Gemini 3 Pro ships as a **preview**, meaning a version open to everyone but still being refined. Here is where it appeared:

| Who | Where |
|---|---|
| Everyone | The Gemini app |
| Google AI Pro and Ultra subscribers | AI Mode in Google Search |
| Developers | Gemini API in Google AI Studio, Gemini CLI, Google Antigravity |
| Businesses | Vertex AI and Gemini Enterprise |
| Other tools | Cursor, GitHub, JetBrains, Replit and others |

Google also announced **Gemini 3 Deep Think**, a mode that thinks longer on very hard problems. It is going first to safety testers, with Ultra subscribers next, within a few weeks.

The scores Google highlights, all its own measurements:

| Test | What it measures | Gemini 3 Pro |
|---|---|---|
| LMArena | People's votes between anonymous models | 1501 Elo (first place) |
| Humanity's Last Exam | Expert-level questions, no tools | 37.5% |
| GPQA Diamond | Graduate-level science | 91.9% |
| SWE-bench Verified | Fixing real bugs in code | 76.2% |
| Terminal-Bench 2.0 | Working through a command line | 54.2% |
| SimpleQA Verified | Short factual questions | 72.1% |

Deep Think raises Humanity's Last Exam to 41.0% and scores 45.1% on ARC-AGI-2, a set of visual puzzles designed to be new to the model.

## How Gemini 3 Pro works

The model card, Google's technical summary of the model, gives the main design choices.

**Built from scratch.** Gemini 3 Pro is not a retrained version of Gemini 2.5; it is a new model, trained on Google's own chips, called TPUs.

**A mixture of experts.** Inside, the model is divided into many specialist sub-networks, or "experts." For each **token**, a small piece of text such as part of a word, a router sends the work to only a few of them. Think of a large hospital where each patient sees the relevant specialists rather than every doctor. This lets the model be very large without every answer costing the full amount of computing.

**Natively multimodal.** It was trained from the start on text, images, audio and video together, rather than having vision added later. That is why Google stresses examples such as reading handwritten recipes, analyzing a video of a pickleball match, or turning a lecture into interactive flashcards.

**A very long memory for one task.** Its **context window**, the amount it can read at once, is up to 1 million tokens, enough for several books or a large code project. Its answers can run up to 64,000 tokens.

The model card also lists limits: like other models it can **hallucinate**, meaning state false things confidently, and its training knowledge stops in January 2025.

## A worked example

Google says Gemini 3 Pro is good at turning dense material into something you can explore, often by writing code for a small interactive page. A request that plays to those strengths:

```prompt
Here is a 40-minute lecture video on how vaccines train the immune system.
Make an interactive study page: a one-paragraph summary, a labeled
diagram of the steps, and ten flashcards that flip to show the answer.
Use plain language for a high-school student.
```

The model watches the video, picks out the key steps, and writes the code for a page with the diagram and flashcards. You can then ask for changes, such as harder questions or a different layout.

> **Try it:** Take a photo of a handwritten note, recipe or whiteboard and ask Gemini to turn it into a clean, organized document. Check whether it read every word correctly.

## Why it matters

Google is competing on reach as much as on test scores. Pichai notes that AI Overviews reach 2 billion people a month and the Gemini app more than 650 million; putting the new model straight into Search puts it in front of far more people than a chatbot alone.

Two other launches point to where Google is heading. **Google Antigravity** is a coding environment where AI agents get their own workspace with access to the editor, the terminal and a browser, so they can plan, build and test software while the developer supervises. **Gemini Agent**, for Ultra subscribers in the Gemini app, handles multi-step chores such as organizing an inbox, with the user kept in control. Google also says the model is less **sycophantic**, less inclined to flatter, and harder to fool with **prompt injection**, hidden instructions planted in content to hijack an AI.

## How can you start using Gemini 3 Pro today?

The simplest route is the Gemini app. Google says Gemini 3 is rolling out there for everyone, so start a chat and give it something it handles well: a document, a photo or a video together with a clear request. If you subscribe to Google AI Pro or Ultra, you can also try it in AI Mode in Search, where answers may come with interactive layouts built for your question.

If you build software, open Google AI Studio and pick the Gemini 3 Pro preview, or try it from the Gemini CLI or Google Antigravity. Start with a small, testable task, such as a single web page, and review what it writes before you trust bigger jobs to it.

Whatever you use it for, treat the benchmark scores as Google's own numbers, not a guarantee. Test it on your real tasks, compare it with the tools you already use, and check important facts, since the model can still make mistakes.

## Sources

- A new era of intelligence with Gemini 3: https://blog.google/products/gemini/gemini-3/
- Gemini 3 Pro - Model Card: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf