Gemini 3 vs 2.5: Battle of the Models Round 2, Fight! | by Angga Agia Wardhana | Google Cloud - Community | Medium
Gemini 3 vs 2.5: Battle of the Models Round 2, Fight!
8 min read
·
Nov 19, 2025
--
1
Share
Disclaimer: This entire comparison is done inside AI Studio, powered by caffeine, curiosity, and zero scientific rigor. Nothing here follows academic standards. Gemini 3 is still in preview, so don’t come at me at 2 AM if your results look different. This exists for fun, vibes, and general capability checking.
If you are like me (who was literally sleeping when this happened), you woke up to find Google casually dropping Gemini 3 in the middle of the night (Jakarta time, spare me).
Gemini 2.5 barely had time to enjoy the spotlight. It literally just got here. And suddenly, Gemini 3 slides in like a new DLC boss character with better stats, cooler animations, and a mysterious power level that Google refuses to fully explain.
Which means only one thing for us. Time to run it through the same gauntlet as last time.
Welcome back to Round 2 of “casually stress testing multimodal models so I have a new Medium article material.”
Let’s jump into the fights. Ding 🔔
The Setup
I’m using Google AI Studio again. Quick tip: they have this feature built-in now. Here it is:
after that you can choose which model you want to compare:
Press enter or click to view image in full size
Testing parameters
- Left Corner: Gemini 3 Pro Preview (Temp: 1, Thinking: HIGH)
- Right Corner: Gemini 2.5 Pro (Temp: 1, Thinking Budget: 8,000 tokens)
I tried to keep it fair, but Gemini 3 uses a dynamic “thinking level” while 2.5 uses a hard token budget. We work with what we have.
Round 1: Logic Puzzle (No Math Tricks)
The Prompt:
A farmer has 17 cows. All but 9 die. How many are left?
Explain your reasoning clearly but without revealing your step by step internal chain of thought.
Result:
Press enter or click to view image in full size
Both models got the answer right (9). No surprises there. But the vibe difference? Massive.
- Gemini 3 Pro ✅: Answered instantly. Clean, compact, efficient. Even with “High Thinking” on, it didn’t overthink. It just dropped the mic.
- Gemini 2.5 Pro ⚠️: It took its sweet time. It spent way more thinking tokens just to arrive at the same conclusion. The answer was right, but the process felt bulky.
Verdict:Gemini 3 Pro wins Round 1 for speed and not wasting my time.
Round 2: The Summarization Challenge
System instructions
Your task is to summarize text clearly and concisely while keeping the author's tone, vibe, and intention. Do not rewrite the text into a formal article. Maintain the emotional context.
Prompt:
Summarize the following text into 5 bullet points while keeping the original tone and emotional weight. Do not oversimplify or remove important nuance.
TEXT:
"So earlier midnight while I'm sleeping, Google casually drops Gemini 3. Gemini 2.5 barely had time to enjoy the spotlight, and suddenly Gemini 3 slides in like a new DLC boss character with better stats, cooler animations, and a mysterious power level that Google keeps hinting at.
Which means only one thing for us. Time to run it through the same gauntlet as last time.
Welcome back to Round 2 of 'casually stress testing multimodal models so I have new Medium article.'
Let's jump into the fights."
Result
Press enter or click to view image in full size
This is where things got wild.
- Gemini 3 Pro ✅: It understood the assignment. It kept the “casual chaos,” the DLC boss joke, and the “welp here we go again” energy. It didn’t just summarize; it vibe-coded the output.
- Stats: ~312 tokens, ~10s thinking.
- Gemini 2.5 Pro ⚠️: It gave me the “Corporate HR” version. It summarized the facts correctly but stripped out all the flavor. It felt flat.
- Stats: ~2,144 tokens, ~17s thinking.
Verdict:Gemini 3 Pro wins. It’s not just smarter; it has a personality. Plus, a 7x difference in token usage for a better result? That’s illegal.
Round 3: Creative Writing (Short Story)
Task: Create a short fictional story (3–4 sentences) with a clear setting, character, conflict, and a surprising or emotional twist.
System instructions:
You are a creative writing AI. Write with vivid imagery, emotional clarity, and concise storytelling.
Prompt:
Write a short fictional story (3–4 sentences). The story must: introduce a character and setting quickly show a conflict or dilemma end with a twist or emotional punch keep the tone imaginative and engaging Do NOT exceed 4 sentences. Begin.
Result:
Press enter or click to view image in full size
- Gemini 3 Pro ✅: Wrote a tight, emotional story about a “Star-Weaver” in just 188 tokens. It hit the emotional punch without rambling.
- Gemini 2.5 Pro ⚠️: Wrote a novella. It wasn’t wrong, it was just… a lot. It used 1,464 tokens to do what Gemini 3 did in 188.
Verdict:Gemini 3 Pro. Brevity is the soul of wit (and lower API bills).
Round 4: Climate Change Explanation
Prompt:
Explain climate change in the simplest way possible so that anyone — including someone with zero science background — can clearly understand it. The explanation must be:
short (4–6 sentences)
clear and accessible
free of jargon
focused on the core idea, not debates Begin.
Result:
Press enter or click to view image in full size
- Gemini 3 Pro ✅: “Earth is wrapped in a blanket.” Simple. Clean. fast. Zero unnecessary fluff.
- Gemini 2.5 Pro ⚠️: Still good, but heavy. It used a massive thinking budget to produce a textbook explanation that felt slower and denser.
Verdict:Gemini 3 Pro wins on clarity and speed.
Round 5 — Python Coding Challenge
Prompt:
Write a Python function that takes a list of numbers and returns the sum. Keep the code minimal.
Result:
Press enter or click to view image in full size
- Gemini 3 Pro ✅:
def sum_list(numbers): return sum(numbers)
used 42 tokens. It gave me exactly what I asked for. Minimal. Perfect.
- Gemini 2.5 Pro ⚠️: Burned 901 tokens. It gave me the code, then added a docstring, then added example usage, then explained how the code works. Bro, I just asked for the function.
Verdict:Gemini 3 Pro. It listens to instructions. Gemini 2.5 is that intern who tries too hard.
Round 6: The “Explain Like I’m 5 to PhD” Test
Prompt:
The Prompt: Explain photosynthesis to four audiences:
A 5-year-old
A high school student
An undergraduate biology major
A PhD candidate in plant sciences.
Result
Press enter or click to view image in full size
- Gemini 3 Pro ✅: Used 1,197 tokens. The transitions were smooth. The PhD section focused on regulation and research frontiers (the actual smart stuff) without drowning in text.
- Gemini 2.5 Pro ⚠️: Used 4,827 tokens. That is 4x more. It gave a strong academic dump, but it lacked the “reading the room” ability that G3 showed.
Verdict:Gemini 3 Pro. It knows when to expand and when to compress.
Round 7 — Phishing Email Detection
The Prompt
Analyze the following description of a potential phishing email image and identify key elements that might indicate a phishing attempt.
Result
Press enter or click to view image in full size
- Gemini 3 Pro ✅: Acted like a digital detective. It caught subtle cues like punctuation errors and domain spoofing strategy. It was structured, forensic, and used half the tokens of 2.5.
- Gemini 2.5 Pro ⚠️: Solid, but generic. It felt less like an incident response report and more like a general warning.
Verdict:Gemini 3 Pro. Sharper eyes, faster report.
Final Scoreboard 🏁
Press enter or click to view image in full size
🥇 The Winner: Gemini 3 Pro
(if you look at the table, basically gemini 3 dusts 2.5 by massive margin)
Press enter or click to view image in full size
Food For Thought 🍔
1. Efficiency is the new King. Gemini 3 used 73.9% fewer tokens to do the exact same work. If you are paying for API calls, this is money in your pocket.
2. Speed Kills. Gemini 3 finished the gauntlet in half the time. It doesn’t “brute force” reasoning anymore; it just gets it.
3. The “Vibe” Upgrade. Gemini 2.5 often felt like a “corporate rewrite”: safe, wordy, and flat. Gemini 3 feels like it has personality. It understands nuance, humor, and brevity.
Bottom Line: If this were an actual boxing match, Gemini 3 didn’t just win on points, it lapped 2.5 around the ring while throwing fewer punches.
Want to nerd out on Gemini 3?
If you want to see the actual science (or just check if I made this all up), check the official drops:
- The Model: deepmind.google/models/gemini/
- The Announcement: blog.google/products/gemini/gemini-3
- The Collection: blog.google/products/gemini/gemini-3-collection/
The “Insider” Insight
After digging through the docs (so you don’t have to), here is why Gemini 3 crushed the tests:
- It’s Literally Built for “Vibes”: Google describes Gemini 3 as their best “vibe coding” model. That isn’t just marketing fluff — it explains why it nailed the tone in Round 2 while Gemini 2.5 sounded like a corporate HR memo. It understands intent , not just instructions.
- Thinking Level > Thinking Budget: Gemini 2.5 relied on a “token budget” (like giving a kid an allowance). Gemini 3 uses a dynamic “Thinking Level.” It thinks exactly as long as it needs to, then executes. That is why it was 2x faster — it doesn’t waste time pretending to be smart; it just is.
- The Robots Are Getting Hands: This release launched alongside Google Antigravity , a new platform for AI agents. If Gemini 3 is the brain, Antigravity is the hands. Basically, the AI isn’t just reading your code anymore; it’s preparing to write the whole app for you.
Welcome to the new era. Now, if you’ll excuse me, I’m going back to sleep. 😴
P.S Google Cloud credits are provided for this project.
About the Author
I’m an IT Manager by day, a GDE by night, and just a simple man with a chaotic curiosity for AI and Cloud. I write about tech with zero academic rigor, 100% vibes, and arguably too much caffeine. If you enjoyed this fight, smash that clap button until your mouse breaks or Medium bans you. 👏
