Jeya Shad

Posted on Jan 10 • Originally published at generativeai.pub on Jan 10

How to Use Gemini 3: The New "Deep Think" Mode Explained (Plus the "Antigravity" Feature No One Is Talking About)

#googleantigravity #antigravity #ai #googlegemini3

Quick Note: if you want to see something cool?? Reach the Hidden Tricks Section!!!

1. Introduction: The AI Frontier Just Got a Major Upgrade

Google has just released Gemini 3, and trust me, the impact is huge. This is not a small update or a quick bug fix. No, friends, this is Google’s smartest, most capable model to date.

We’re talking about an AI that can understand text, images, audio, and more all at once. It can write and fix code so well it almost feels unreal. And with its growing abilities, it’s easy to imagine a future where this AI might even book your next vacation for you. (hopefully to somewhere more interesting than your in-laws’!).

Just to clear up any confusion, we’re talking about Google’s AI here, not NASA’s old Gemini 3 space mission. But honestly, both have something in common. They both try to push limits. One explored space. The other explores the fast-growing world of artificial intelligence.

https://x.com/sundarpichai/status/1990812770762215649

I haven’t slept much since the announcement yesterday.

Usually, when a new feature or model comes out, I always plan to write a blog about it, but I never actually get around to doing it. I read the announcement, test a few prompts, and move on with my day. But Gemini 3 is different. This time, I felt like I had to write about it. It’s the first time I’ve felt like the AI isn’t just answering me; it’s actually working with me.

If you’re feeling overwhelmed by the new updates, don’t worry. I’ve spent the last 24 hours breaking it down. Here is everything you need to know, from the headline features to the stuff hidden in the developer settings.

2. It Doesn’t Just Talk. It “Acts.”

The headline feature of Gemini 3 is Agency.

Previously, if you asked AI to “plan a vacation,” it would give you a text list of hotels. You still had to go to the websites, check dates, and book them. With Gemini 3, we have entered the “Agent Era.”

You can say:

“Find a flight to London under $600 for next Tuesday, put it on hold, and block that time on my calendar.”

And it just… does it. It navigates the web, filters the results, and executes the task. It’s no longer just generating text; it is generating actions.

3. “Deep Think” Mode: The Pause That Changes Everything

The first thing you’ll notice in the app is a new toggle next to the send button: Deep Think.

Use Deep Think in Gemini Apps

This is the real game-changer. In older versions like Gemini 1.5 or 2.0, the AI focused on answering as fast as it could, and that speed often led to confident mistakes.

Even though Gemini 2.5 introduced “thinking” mode and improved accuracy, this new release goes far beyond that. Gemini 3 doesn’t just think better; it pushes past old limits, breaks barriers, and feels like a truly upgraded version of what AI can be.

When you toggle Deep Think on, Gemini 3 slows down.

The Experience: You ask a hard question. Instead of instant text, you see a pulsing “Thinking…” indicator. It simulates a chain of thought checking its logic, spotting errors, and re-planning its answer before it types a single word.
My Test: I gave it a complex logic puzzle that Gemini 2.5 failed every time. Gemini 3 paused for 12 seconds, “thought” through the trap, and gave me the correct answer.
When to use it: Coding, legal questions, or math.
When to skip it: Asking for a playlist or a recipe. You don’t need PhD-level reasoning to boil pasta.

4. Under the Hood: The Brains Behind the Brilliance

Alright, let’s get technical for a moment. What is it, precisely, that makes Gemini 3 tick? What sorcery lies beneath the surface, enabling it to achieve such impressive feats?

At its heart, we find the “Dynamic Routing architecture,” a sophisticated Mixture-of-Experts (MoE) system. Think of it as a brain with many specialized modules, each honed for a particular task. This allows Gemini 3 to be incredibly efficient, focusing its computational resources where they’re needed most (like coding or math), while simultaneously maintaining a broad base of knowledge.

The “Deep Think” Mode

And then there’s “Deep Think Mode,” the AI equivalent of donning your most formidable thinking cap. This is where Gemini 3 truly shines, decomposing complex problems into manageable sub problems, evaluating different solution paths, and even self correcting along the way.

It’s this capacity for “PhD-level reasoning” that allows it to tackle challenges that would leave lesser models scratching their digital heads.

The Experience: When you use this, you see a pulsing “Thinking…” indicator. It takes 10-20 seconds, but the accuracy is unmatched.
The Stats: In the GPQA Diamond benchmark , it scored 93.8% a score most human experts can’t touch.

Let’s not forget the context window, now expanded beyond 2 million tokens. We’re talking about processing entire sagas, not just fleeting snippets. This is crucial for maintaining coherence in extended sessions, allowing Gemini 3 to remember and build upon previous interactions.

5. Google Antigravity: The “Secret” Feature for Builders

Okay, this wasn’t in the main Super Bowl-style ad, but it is easily the most insane part of the release for developers. Google launched a platform called Antigravity.

Key Resources

Let’s start with a few resources that give you a quick reference to the official documentation and materials for Antigravity (as of November 19, 2025):

Official Site: https://antigravity.google/
Documentation: https://antigravity.google/docs
Use Cases: https://antigravity.google/use-cases
Download: https://antigravity.google/download
YouTube Channel: https://www.youtube.com/@googleantigravity

These links provide everything you need to explore Antigravity in detail, from official guides to hands-on tutorials and examples.

Think of it as giving the AI “hands.” It integrates directly into tools developers already use, like VS Code, JetBrains, Cursor and Android Studio. Until now, coding with AI meant copy-pasting code from a chat window into your editor. Antigravity changes the workflow entirely. It gives Gemini 3 direct access to a code editor, a terminal, and a browser.

Manager View: I opened Antigravity this morning, and it felt like I was a manager supervising an intern. I didn’t write code; I just approved plans.
Artifacts: This is a huge quality-of-life upgrade. Instead of a wall of text, Antigravity generates “Artifacts” clean, visual cards showing the code it wrote, screenshots of the bugs it found, and a checklist of what it’s doing next.
The “Holy Cow” Moment: I asked it to “fix the spacing issue on my website.” I watched (hands-free) as it opened the file, wrote the fix, opened a virtual browser, took a screenshot to verify it worked, and then asked me for approval.

6. Vibe Coding & Generative UI (The Fun Stuff)

If you aren’t a developer, this is for you. The new “Vibe Coding” capability means you can build software just by describing the “vibe.”

I tried the prompt from Twitter:

“Make me a Snake game where the snake is a cat and the food is sushi.”

In Gemini 2.0 or 2.5, it would have just spit out Python code that I wouldn’t know how to run. In Gemini 3, it triggered Generative UI. It literally built the game inside the chat window. I was playing “Cat Snake” 10 seconds later. It even added a retro 8-bit soundtrack because I used the word “retro” in my follow-up.

But it’s not just for games. I also tried:

“Help me choose a mortgage.”

Instead of explaining interest rates, Gemini 3 instantly built a custom, interactive calculator right there in the chat window. I could drag sliders and click buttons on a tool that didn’t exist 10 seconds prior.

7. The Scoreboard (For the Data Nerds)

If you care about the hard numbers, Gemini 3 didn’t just beat the competition; it broke the scale.

If you are wondering where I am getting this data, I spent the morning analyzing the LMSYS Chatbot Arena. For those who don’t know, LMSYS is the “gold standard” for AI benchmarking. It isn’t just a static test; it is a blind battle arena where users vote on which AI gives the better answer without knowing which model is which.

(If you want to see exactly how they calculate these rankings, check out their methodology page here. it’s fascinating stuff.)

I dug into their latest November update, and the results are shocking.

I. The WebDev King (Why Coders Are Switching)

I specifically looked at the WebDev Arena , which tests how well an AI can build functional web applications (not just simple snippets). This leaderboard compared 13 top models across 19,621 user votes.

Here is what I found:

Figure 1: The WebDev Rankings Table

Look at that gap. Gemini 3 is nearly 100 points ahead of GPT-5 Medium. In the world of Elo ratings, that isn’t just a “win”!! No, that is a slaughter. It explains why the “Vibe Coding” feature feels so magical; the model understands full-stack structure better than anything else on the market.

II. The Overall Score (A Clean Sweep)

But it’s not just coding. When you zoom out to the Overall Leaderboard , Gemini 3 has effectively swept the board. According to the latest snapshot, it holds the #1 Spot in almost every single category:

Figure 2: LMSYS Chatbot Arena Overview showing the performance rankings of the most recent AI models.

The closest competitor is Grok-4.1-Thinking , which sits at 1484 Elo. Even the previous champion, Gemini 2.5, has been pushed down to rank #3.

Figure 3: The Scoreboard: The New King of the Arena

For the last three months, Gemini 2.5 Pro and Claude 4.5 have been trading blows for the #1 spot on the LMSYS Chatbot Arena. Yesterday, Gemini 3 cleared them both by a 50-point margin-a gap we haven’t seen since GPT-4 first launched.

III. The “Generational Leap” Comparison

To visualize just how big of a jump this is from the version you were likely using last year (Gemini 1.5), look at the specific benchmark comparisons:

Figure 4: The Generational Leap Table (Gemini 1.5 vs Gemini 3)

If you want to see the full breakdown or test the models yourself, I highly recommend going to lmarena.ai and checking the “Leaderboard” tab.

Finally, The headline scores are, to put it mildly, impressive:

LMArena Leaderboard: 1501 Elo - a new high score!
Humanity’s Last Exam: 37.5% (41.0% with Deep Think mode) without tools.
GPQA Diamond: 91.9% (93.8% with Deep Think mode).
MMMU-Pro (81%) and Video-MMMU (87.6%) for multimodal reasoning.
MathArena Apex: 23.4% - setting a new math standard.
SWE-bench Verified (76.2%) and Terminal-Bench 2.0 (54.2%) for coding and tool use.

8. The Buzz & The Bumps: Current Takes & Past Troubles

The response to Gemini 3 has been extremely positive. Developers and early users are already calling it “far more capable,” especially praising how well it writes code and how surprisingly useful its answers are.

What’s interesting is that Google didn’t make a huge splash with this launch. Compared to past releases, the announcement felt quieter and more careful. Some people think this is on purpose, almost like a “perfection first” strategy, especially after the issues they faced before with image generation mistakes and unstable APIs.

Still, a few concerns are hanging around. One big one is the “Enterprise Squeeze,” where businesses struggled because older models were removed without clear upgrade paths. Many are now watching to see if Google handles things better this time.

https://gemini.google/release-notes/

9. Hidden Tricks I Found (That Google Didn’t Mention)

After poking around for hours, here are three things I found that you should try:

The “Critique” Button: After Gemini gives you an answer, hover over it. There’s a tiny magnifying glass icon. Click it, and Gemini will re-read its own response to find holes in its argument. It’s brutal but useful.
Video Diagnostics: I pointed my camera at my Nespresso machine which was blinking red. I didn’t say anything. Gemini 3 watched the video, heard the specific whirring sound, and said, “That sound means the water pump is airlocked. Here is how to fix it…”
The “Manager” Prompt: If you are in Antigravity, try telling it: “Run in autonomous mode for 3 steps.” It will try to solve a problem for three consecutive steps without asking you for help. Scary good.
The “Time Machine” Trick (My Favorite): This sounds impossible, but try it: Choose Canvas and Ask Gemini to “Build and render a fully functional interactive mockup of Windows 95” Because of the new Antigravity engine, it doesn’t just describe the OS; it writes the full HTML/CSS/JS bundle and renders it as an interactive “Artifact” in the chat. You can actually click “Start,” drag windows , and use the menus. It’s a wild demonstration of its coding speed.

Google Gemini 3 canvas

Trigger “Dynamic View” (The App Builder)

This is the technical name for the new Generative UI feature, but you have to know how to trigger it. If you just ask “tell me about mortgages,” you get text. But if you say “help me understand DDoS 🤓” Gemini 3 enters Dynamic View. It stops acting like a chatbot and starts acting like a software engineer, building a bespoke, interactive interface just for you.

Gemini 3 Dynamic View (Lab)

Here, I’m sharing a prompt I typed and the result it generated: Try it yourself. You can experience it too.

I really love this feature because it makes learning not just easier, but also fun. It helps me understand things in a way I’m more likely to remember. I have to thank Google for this, as a growing software engineer, tools like Gemini 3 make studying complex system architectures much more approachable and enjoyable.

10. Final Verdict: The Agent Era is Here

I am usually skeptical of “revolutionary” updates. We’ve heard that word too many times.

But Gemini 3 is the first time I’ve felt like the AI isn’t just a tool I use, but a teammate I trust.

It doesn’t just give travel advice; it books the flight.
It doesn’t just explain code; it builds the app.
It doesn’t just guess; it thinks.

If you are still using the old models, you are effectively working with one hand tied behind your back. The upgrade is free, the features are live, and the “Chatbot Era” is officially over.

Gemini 3 is rolling out now. Have you tried the “Time Machine” prompt yet? Let me know in the comments if you managed to get Doom running on it.