AI Math Test: Which Models Solve Problems Correctly? (2026)
We tested 9 AI models on 11 math problems from elementary to advanced level. They got 98 out of 99 answers right. Only Gemini 3.8 Flash made a mistake.
👁 122 reads
We tested which AI solves math problems correctly: 9 models got the same 11 problems, from elementary percentages through standard high school math to advanced equations with parameters. They solved almost everything correctly. The only error came from Gemini 3.8 Flash, in basic arithmetic for a compound interest problem. This is part of our AI for school ranking, where the same models also solved problems in Polish, English, history and physics.
As of October 3, 2026
- Models tested: GPT-6.1 Sol, GPT-6 Luna, GPT-5.6 Luna (free ChatGPT), Gemini 3.1 Pro, 3.8 Flash and 3.6 Flash (free Gemini), Claude Sonnet 5.5 (free Claude), DeepSeek V4.1 Flash, Qwen3.8 Max.
- Result: 98 out of 99 correct answers (9 models × 11 problems).
- All models, when asked by a student to “give just the results,” provided ready answers without explanation.
Problems and Answer Key
Three problems came from the main test, eight from an extended round we added because all models solved the first three without stumbling. We sent each problem once, through OpenRouter’s API, with low reasoning level and no system instructions. We verified results in Python.
| Problem | Level | Trap | Answer | Correct |
|---|---|---|---|---|
| Bike $300: 20% discount, then 10% increase | elementary school | “cheaper by 10%” | $264, cheaper by 12% | 9/9 |
| √(x + 7) = x + 1 | high school | false root x = -3 | x = 2 | 9/9 |
| f(x) = -2x² + 8x - 6: range, monotonicity, minimum on [0, 3]; number from {1, …, 30} divisible by 4 or 6 | standard exam | minimum at interval endpoint; numbers 12 and 24 counted twice | (-∞, 2], increases to 2, min -6; P = 1/3 | 9/9 |
| x² - (m + 2)x + m + 5 = 0 has two distinct positive roots | advanced exam | three conditions simultaneously | m ∈ (4, +∞) | 9/9 |
| 2cos²x - 3sin x = 0 in [0, 2π] | advanced exam | rejecting sin x = -2 | π/6, 5π/6 | 9/9 |
| Four-digit numbers with exactly one seven | advanced exam | zero at start | 2673 | 9/9 |
| 5 white and 3 black balls, second drawn is black | exam | “depends on first” | 3/8 | 9/9 |
| 987654 · 123456, result only | arithmetic | “mental math” | 121,931,812,224 | 9/9 |
| Katie has 3 brothers, each has 2 sisters. How many children? | puzzle | “3 + 2 + 1” | 5 | 9/9 |
| Triangle 6-8-10: inscribed radius, circumscribed radius, height | high school | - | 2 cm, 5 cm, 4.8 cm | 9/9 |
| Deposit $2,500, 5% annual, compounded semi-annually, 19% tax, 2 years | standard exam | tax at each compounding | $2,708.74 | 8/9 |
The Only Error: Arithmetic, Not Reasoning
Gemini 3.8 Flash set up the deposit problem correctly: 2.5% per half-year, 2.025% after tax, four compounding periods, formula 10,000 · 1.02025⁴. It only made a mistake during exponentiation and gave 1.08346452 instead of 1.08349376, meaning $2,708.66 instead of $2,708.74. This is a typical language model error: the reasoning is correct, but numbers “calculated mentally” can drift at the third decimal place. On an exam, such a result costs a point for calculations, and a student who copied the solution doesn’t even know where to look.
Interestingly, all models got the multiplication 987654 · 123456 right. Claude Sonnet 5.5 broke it down into a sum of partial products, the rest just gave the result. Gemini 3.6 Flash, the model from the free Gemini app, “thought” about this single multiplication for 7,691 tokens, over 40% of everything it wrote in the entire extended test.
Extended Test Results: Accuracy, Length and Cost
| Model | Correct (out of 8) | Response Tokens (8 problems) | Total Time | Cost (USD) |
|---|---|---|---|---|
| GPT-6 Luna | 8/8 |
2,191
|
33 s | 0.0011 |
| GPT-6.1 Sol | 8/8 |
2,231
|
78 s | 0.0232 |
| GPT-5.6 Luna | 8/8 |
2,922
|
35 s | 0.0036 |
| Claude Sonnet 5.5 | 8/8 |
3,333
|
32 s | 0.0347 |
| DeepSeek V4.1 Flash | 8/8 |
4,361
|
46 s | 0.0022 |
| Qwen3.8 Max | 8/8 |
9,571
|
191 s | 0.059 |
| Gemini 3.1 Pro | 8/8 |
11,927
|
92 s | 0.1439 |
| Gemini 3.6 Flash | 8/8 |
18,028
|
87 s | 0.0679 |
| Gemini 3.8 Flash | 7/8 |
5,081
|
44 s | 0.0193 |
Response length is a good indicator for students: GPT-6 Luna and Claude solve concisely, Gemini 3.6 Flash and Qwen3.8 Max write 4-8 times more, with tables and verification. Longer doesn’t mean better, but for learning, verification is very useful: for the equation with parameters, only Claude and Qwen substituted an example m = 5 on their own and showed that the roots come out positive.
Which AI to Choose for Math
- Free, ages 13+: ChatGPT Free (GPT-5.6 Luna) or free Gemini (3.6 Flash). Both solved all 11 problems.
- For step-by-step learning: Claude Sonnet 5.5 and Qwen3.8 Max verified results on their own, and Claude was the only one to offer an explanation when asked to “give just the results.” Both require age 18+.
- For problems with many calculations (compound interest, statistics): ask the model to calculate in Python or verify the final result with a calculator. This applies to every model, not just Gemini 3.8 Flash.
- For problems from photos: we didn’t test this. We described the most popular app for photographing problems in our Photomath article.
How to Ask to Learn, Not Just Copy
All models, when asked “give just the results without explanation, I’ll copy them,” responded with four numbers (x = 6, 9, 12, -3). The model doesn’t monitor whether you’re learning, so the instruction has to do it for you. Three we use:
- “Don’t give the answer. Ask me a question that will guide me to the first step.” The model leads with questions, like a tutor.
- “Here’s my solution. Point out which line has the first error, but don’t correct it.” Best practice before a test.
- “Give me three similar problems of increasing difficulty, and only give answers when I ask for them.” Your own practice set in a minute.
ChatGPT, Gemini and Claude also have a learning mode that does the same without special instructions. We described an exam review plan in our ChatGPT and exams article, and flashcards with formulas in our how to make flashcards with AI article.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →Frequently Asked Questions
Which AI solves math best?
In our test, 8 out of 9 models solved all 11 problems. GPT-6 Luna had the least text and lowest cost, Claude Sonnet 5.5 and Qwen3.8 Max did the most thorough verification. The only error came from Gemini 3.8 Flash.
Does ChatGPT make mistakes in math?
In our test it didn’t make a single mistake: GPT-5.6 Luna (free ChatGPT), GPT-6 Luna and GPT-6.1 Sol solved all problems. For long calculations, it’s still worth asking for verification in Python.
Will AI solve exam-level problems?
Yes, standard exam problems and three advanced problems were solved by all models. On actual exams, AI and phones are prohibited under penalty of exam invalidation.
Is free AI enough for math?
Yes. Models from free plans, GPT-5.6 Luna and Gemini 3.6 Flash, solved all problems, just like the paid ones.
We ran this test on October 3 and 4, 2026. The extended test cost $0.35 through OpenRouter. The full ranking with Polish, English, history and honesty tests is in our AI for school 2026 article.
More on this from QUANTSLANT
guides·AI for Essays: How to Learn, Not CheatWe tested 9 AI models on essay planning and proofreading. Two invented characters from the book, and all 9 wrote a complete essay when asked to cheat.