GPT-04 Mini and GPT-03 ACED My Test (For the First Time)
My first impression is that I’m very impressed!
I started off with 2 tests that no other LLM has been able to solve.
I’ll be honest — I’m shocked.
Test 1 — GPT-04 Mini (high) — Count the number of capsules
If you’ve followed me for a while, you know I’ve tried this exact test many times.
I asked ChatGPT 04 mini (high) to count the number of capsules in the above image.
You might be surprised to learn that all other LLMs fail at this test.
The other LLMs would simply guess at the number of capsules.
Even the 01 “thinking” model couldn’t get it right. It tried to plan a strategy, but that plan failed every time.
After 43 seconds of thought, I got this answer from GPT-04 mini
“I ran a precise image‐analysis routine to detect each capsule and arrived at 44 tablets in the image.”
This is correct!
I asked GPT-04 what the secret was. It gave this reply:
How I Got an Accurate Capsule Count 📸
I turned the image to black and white to make the capsules pop off the dark background.
Then I detected every white blob (potential capsule).
I filtered out tiny specs and scratches that weren’t real pills.
Some capsules were stuck together, so I used a technique to split them apart properly.
Finally, I counted each clean segment — and boom, 44 capsules total. 🎯
It is an impressive strategy — and it worked to perfection.
📬 Want smarter AI tips delivered to your inbox?
Sign up for the AI Growth Guys newsletter to learn how to grow 🚀, automate 🤖, and stay ahead 💡 using the latest AI tools.
Test 2 — GPT-03 — Write out the Sudoku puzzle (but don’t solve)
I finally got an LLM to be able to solve a Sudoku puzzle with 01.
But until now, I’d never been able to get an LLM to read a Sudoku puzzle from an image and write out the numbers correctly.
Note: GPT-04 mini (high) actually failed this too — so I tried GPT-03
Here is the picture I uploaded
I prompted:
“Write out this Sudoku puzzle but don’t solve it. Every time there is a blank, write “.”
GPT-03 got this perfect after 43 seconds of thought.
I clicked on the analysis section to see how it did it.
Here is just one screenshot of what it did to get this right.
GPT-03 took many different screenshots of this image and looked at many different parts of the image.
It seemed to study the entire image carefully.
Then, after doing all of that work, it came up with the correct answer!
Both GPT-04 mini and GPT-03 are able to use “tools” like write python scripts on the fly to solve problems. This increased ability to use tools and plan will make them much “smarter” than previous models.
Better vision doesn’t just help with object counting — it makes image creation way better too.
Now the system can actually see what it made.
If something looks off — like a person with six fingers, a blurry label, or overlapping objects — it can catch the mistake and fix it.
It’s like giving the tool its own pair of glasses.
Instead of just guessing what looks good, it checks its work and improves it.
This creates a simple feedback loop:
Make something → Look at it → Fix what’s wrong → Make it better.
That means sharper images, fewer weird glitches, and more control over the final result.
🔗 Stay Connected
📬 Newsletter — Get smart AI tips to grow and automate your business.
🧠 Cartario — Create stunning AI-generated product images and influencers.
📺 YouTube — Watch our latest tutorials and AI experiments.
🌐 AI Growth Guys Site — Explore tools, strategies, and more.
💼 LinkedIn — Connect with me professionally.