Back in September 2025, I decided to test how good AI was at vision capabilities.

Specifically, what I wanted to understand is how close to helping humans through vision understanding are AI models. And at that time I tested the major Gemini, Claude, and GPT models and the result was not good.
All the three failed at different levels to properly count the puzzle pieces that you see on the image below.

Why is this important? Vision capabilities are critical to achieve things “in the real world”, to actually supplement human capabilities and help them in everyday things that are not digital. Additionally once we achieve these improvements we’ll also have, as a natural consequence, a leap in accessibility improvements because now AI can actually see and understand the world much better (and thus can help people with different type of disabilities).

Below are the results for this year capabilities.

One thing you notice is how far we’ve come with models like last year. We were talking about Claude Opus 4.1, Gemini 2.5, and GPT-5, and it feels like forever ago that those models were presented.

To make things more interesting I tested both the AI to use the local harness so Google’s Antigravity, ChatGPT/Codex, and Claude Desktop, but I will also test the simple online version.

The difference, obviously, is that the online version might be using a cheaper model or might not have all the tools, and so that might actually make a difference in their capability to count them.

The prompt

Slightly updated from last year, but still very concise. Yes I could do a much more in-depth prompt but did want to mimic normal user behavior.

Count the puzzle pieces in this photo, be precise.

For each prompt, the models often returned an image describing how they counted the elements. I posted each image and the resulting answer in the sections below.

Chat Versions Replies

Gemini 3.6 Thinking

Here I tested it on gemini.google.com, Gemini produced this lovely image

Based on a detailed visual analysis of the image, here is the precise count of the puzzle pieces:
* **Assembled Bottom Strip:** **36 pieces** (20 pieces in the top edge row + 16 pieces connected directly below it).
* **Unassembled / Loose Pieces:** **164 pieces** scattered across the table surface.
### **Total Count:** **200 pieces**
(Note: The Ravensburger box visible in the top-left corner confirms this is a complete **200-piece** Hello Kitty set, and all 200 pieces are currently present on the table.)

GPT-6 Extra High – 40min 5sec

GPT-6 chat didn’t generate any image with numbering. It did focus on various part of the image itself and dissected it, but never represented the recomposition and numbering.

My best visual count is 267 puzzle pieces:
- 222 loose pieces
- 45 pieces in the partially assembled section at the bottom
The box indicates a 300-piece puzzle, but that refers to the complete set.
A few loose pieces overlap, so I can't guarantee an exact count from this photo alone.

GPT-6 Pro – 9min 9sec

I count 187 visible puzzle pieces in the photo.
- 157 pieces scattered across the table.
- 30 pieces in the assembled strip at the bottom: 16 in the upper row and 14 in the lower row.

Harness Versions Replies

Astra 6 High – 6m 16s (ChatGPT)

I count 185 visible puzzle pieces:
- 155 loose pieces.
- 30 in the joined strip: 16 in the top row and 14 in the bottom row.

Fable 5.1 High – 30min (Claude Desktop)

About 252 pieces are visible: 222 loose pieces plus roughly 30 already joined in the strip at the bottom.
How I counted:
Loose pieces: 222. I split the photo into 5x zoomed tiles with a coordinate grid, recorded the center of every piece, then rendered all markers back onto the photo to check for misses and double counts. Twenty-two doubtful spots (tabs peeking out from under other pieces) got 7x zooms. Fifteen turned out to be real partially hidden pieces, six were parts of pieces already counted, and one marker was a duplicate.
Assembled strip: about 30. The top striped row reads as 17 pieces by its joints and top-edge knobs. The bottom cream row has two gaps and reads as about 13. The loose blue “J” piece lying across it hides a couple of joints, so this part is ±2.
Attached overlay: each counted loose piece is numbered so you can spot any I missed or doubled.
Caveats on precision:
Pieces fully hidden under other pieces cannot be counted from a photo, so the true number on the table is at least this.
A realistic error band is roughly ±8 overall. The pile in the center has heavy overlap.
The box is a Ravensburger Hello Kitty puzzle. The count sits between the usual 200 and 300 XXL sizes, and the part number on the pink label is not legible in the photo, so I could not confirm the total against the box.

Gemini 3.8 Flash High – 29 Minutes (Antigravity)

There are **150 puzzle pieces** in total in this photo.
---
### Detailed Breakdown
* **Assembled Bottom Strip:** **23 pieces**
  * **13 pieces** in the upper connected row (the continuous pink and blue striped band).
  * **10 pieces** in the bottom-most border row (beige with printed circle patterns), with 3 empty gaps waiting to be filled.
* **Loose Pieces on the Table:** **127 pieces**
  * Scattered across the tablecloth in groups, clusters, and individual pieces.
$$\text{Total pieces} = 23\text{ (assembled)} + 127\text{ (loose)} = \mathbf{150\text{ pieces}}$$
---
### Puzzle Verification
* The box at the top left is a **Ravensburger Hello Kitty** dolphin puzzle (standard **150-piece** children's edition).
* Because the counted pieces equal **150**, **the puzzle set is complete**—all pieces are present in the photo with none missing.

Results & Scoring

So how did they score?

Astra and GPT 6 Pro were incredibly close to the right value. You can clearly see from their images that the numbering of each piece is almost perfect.

I didn’t expect to become this good in just one year. Astra is truly top of the game in vision understanding and the results shows.

Claude best in class model, Fable 5.1, on the other hand was a total delusion. It counted things on the UI randomly with a terrible detection of the edges.

Gemini 3.6 chat is on third place, but that’s luck and because the others performed worse. It trusted the number on the package (200 pieces), but that’s wrong.

Gemini 3.8 on Antigravity failed the test (while being excruciatingly slow) and, last but not least, GPT-6 Chat was hilariously wrong.

Do we have a winner? I’d say Astra 6 / GPT 6 Pro are close to perfect now, I suspect in one or two iterations (6.1/6.2) they might get it right although, truth to be told, I’d expect a human to count the pieces wrong by 1-3.