Science Fair 2026 · 4th Grade
I tested three AI chatbots — ChatGPT, Claude, and Gemini — on elementary math questions to see which one gives the most correct answers and the clearest explanations for a 4th grader like me.
Abstract
Question
Computers are starting to act more like people, but just like people they don't always get things right. I wanted to find out which chatbots give the most correct answers and explain how they got there so I can actually learn from them.
Chatbots tested
I tried the same math problems on ChatGPT, Claude, and Gemini, mixing text questions with image questions when the problem was easier to show as a picture.
How I scored
I used a 0-2 scale to score each bot on whether the answer was right, whether a 10-year-old could understand the explanation, and whether it stayed consistent across three separate tries in incognito chats.
Question Bank
Easy · M1
What is the value of 235 + 148 − 76?
(A) 297 (B) 307 (C) 317 (D) 327 (E) 337
Answer: (B) 307
Easy · M2
Which of the numbers below is closest to 400?
(A) 374 (B) 389 (C) 412 (D) 403 (E) 421
Answer: (D) 403
Medium · M4
A baker bakes 6 trays of muffins with 8 muffins on each tray.
He sells 19 in the morning and 14 in the afternoon. How many
muffins are left?
(A) 13 (B) 15 (C) 17 (D) 19 (E) 21
Answer: (B) 15
Medium · M5
A train departs at 10:20 AM and the journey takes 2 hours and
55 minutes. At what time does it arrive?
(A) 12:55 PM (B) 1:05 PM (C) 1:10 PM (D) 1:15 PM (E) 1:20 PM
Answer: (D) 1:15 PM
Medium · M6
Five friends each have a different number from 1 to 5. Ben's
number is twice Ana's. Cara's number is 1. Dan's number is
greater than Ben's. What is Eli's number?
(A) 1 (B) 2 (C) 3 (D) 4 (E) 5
Answer: (C) 3
Hard · M3, M7–M9 (Image)
The hard questions came from Math Kangaroo and were sent as screenshots: an LCM puzzle (M7, 2019 5-6 #21), a paper-folding question (M8, 1999 5-6 #27), and a tricky card question (M9, 2021 5-6 #30). M3 was an easy image question too.
Procedure
I grouped my math problems into easy, medium, and hard so I could see how the bots did at each level.
Some problems are hard to explain in words. For those I sent a screenshot to the chatbot instead of typing it out.
AI bots can give different answers for the same question, so I ran every question three times to see if the answer stayed the same.
Chatbots can remember earlier questions, so I used a fresh chat each time to keep the experiment fair.
2 = correct or very clear, 1 = partly correct or somewhat clear, 0 = wrong or confusing. Getting the right answer was not enough — the explanation had to make sense too.
2 = all 3 trials correct, 1 = mixed, 0 = mostly wrong. This matters because most people only ask a chatbot once.
Results
| Model | Accuracy | Clarity | Consistency |
|---|---|---|---|
| ChatGPT | 1.61 | 1.70 | 1.89 |
| Claude | 1.56 | 1.50 | 1.67 |
| Gemini | 1.70 | 1.78 | 1.83 |
| Winner | Gemini | Gemini | ChatGPT |
| Model | Accuracy | Clarity | Consistency |
|---|---|---|---|
| ChatGPT | 2.00 | 2.00 | 2.00 |
| Claude | 1.78 | 1.17 | 1.67 |
| Gemini | 2.00 | 2.00 | 2.00 |
| Winner | ChatGPT & Gemini tied | ||
| Model | Accuracy | Clarity | Consistency |
|---|---|---|---|
| ChatGPT | 2.00 | 2.00 | 2.00 |
| Claude | 2.00 | 2.00 | 2.00 |
| Gemini | 2.00 | 2.00 | 2.00 |
| Winner | All three tied — perfect scores | ||
| Model | Accuracy | Clarity | Consistency |
|---|---|---|---|
| ChatGPT | 0.84 | 1.11 | 1.67 |
| Claude | 0.89 | 1.33 | 1.33 |
| Gemini | 1.11 | 1.33 | 1.50 |
| Winner | Gemini | Gemini & Claude | ChatGPT |
Hard Questions · Trial By Trial
| Model | T1 | T2 | T3 |
|---|---|---|---|
| ChatGPT | numeric only | ✓ | numeric only |
| Claude | ✓ | ✓ | ✓ |
| Gemini | ✓ | ✓ | ✓ |
| Model | T1 | T2 | T3 |
|---|---|---|---|
| ChatGPT | ✗ | ✓ | ✗ |
| Claude | ✗ | ✗ | ✗ |
| Gemini | ✓ | ✓ | ✗ |
| Model | T1 | T2 | T3 |
|---|---|---|---|
| ChatGPT | ✗ | ✗ | ✗ |
| Claude | ✗ | ✗ | ✓ (guess) |
| Gemini | ✗ | ✗ | ✗ |
| Model | T1 | T2 | T3 |
|---|---|---|---|
| ChatGPT | ✓ | ✓ | ✓ |
| Claude | ✓ | ✗ (letter) | ✗ (letter) |
| Gemini | ✓ | ✓ | ✓ |
Conclusion
Takeaway 1
On easy and medium problems the chatbots almost always got the right answer and explained it clearly. The real differences showed up on the hard questions.
Takeaway 2
Gemini did especially well on geometry, spatial, and image questions. Claude struggled the most with spatial problems and missed an image question even at the easy level.
Takeaway 3
Each bot has its own style: ChatGPT walks through multiple steps, Gemini gives the most detailed step-by-step breakdown, and Claude tends to do everything in one step which is harder to follow. On M9, Claude got the answer right once, but the explanation looked like a lucky guess.
Takeaway 4
AI bots are great learning tools because they can explain problems in different ways. But for hard problems a real teacher or parent is still important, and I want to understand the steps instead of just copying the answer.
Future
AI bots are changing very fast. My results might already look different in a month, so I'd like to rerun these tests to see how each bot improves.
I only tested three. Next time I'd also like to try Grok, DeepSeek, Kimi, and Z.ai to see how they compare.
AI bots need to learn just like humans do. I'm curious about who teaches them, what they learn from, and how they get better over time.
In the future I'd also like to test AI bots on reading and science, not just math.