What prompt makes an AI math solver show its steps?
Paste this into Claude, ChatGPT, Gemini or Whizi and fill the brackets. It makes the model restate the problem, number its steps and check itself.

Solve this [topic, such as algebra or calculus] problem for a [level, such as grade 9 or first year college] student: [paste the whole problem with every number and unit]. First restate what is given and what is asked. Then solve it in numbered steps and name the rule used in each. Finish by substituting your answer back into the problem to check it, and tell me if any step is a guess.
Each part earns its place. The restatement catches a misread problem before it costs you three steps. The named rules let you look any step up. The substitution at the end is the check that catches most arithmetic slips.
Which AI math solver is free?
Most of the solvers Google ranks for this query are free to try, and ChatGPT and Claude have free plans too. The table lists what each tool says on its own page, read on 21 November 2026. A cell that says not stated means the page gave no figure.
| Tool | Free use | Sign up | Paid price | Steps and photo | Best at |
|---|---|---|---|---|---|
| math-solver.io | Says unlimited problems, no cap stated | No account; optional sign in to sync | Not stated | Numbered steps; image or PDF | A fast check with no account |
| MathGPT (math-gpt.org) | Not stated | Not stated | Not stated | Steps; photos and PDFs | Homework from a photo |
| DeepAI Math AI | Not stated | A sign up or login prompt shows | DeepAI Pro $9.99 a month | Steps; up to 10 images a message | Picking among general models |
| NoteGPT | 6 free tries daily, though the page also says unlimited | H1 says no sign up | Not stated | Steps; image or PDF | A few tries with no account |
| Julius | Credit based, amount not stated | Sign in required | Not stated | Steps; photo of a problem | Follow up questions in one chat |
| ChatGPT | Free $0, limited messages with uploads | Not stated | Go $8, Plus $20 a month | Steps when asked; vision limited on Free | Long word problems and code |
| Claude | Free, limited usage every 5 hours and every week | Not stated | Pro $17 a month billed annually, $20 monthly | Steps when asked; image upload not stated | Careful written reasoning |
| Whizi | 10 messages a day on Whizi AI, Auto and Claude Haiku, no card; 5 photos a day | Account needed | Pro $29.99 a month | Steps when asked; attach a photo to a message | Checking one problem on two models |
Whizi loses in two places. math-solver.io and NoteGPT take a problem without an account, and Whizi asks you to make one. A dedicated solver also puts the box on the page, while a chat app expects you to write the prompt above.
Gemini is the other free option. Its help page lists "Standard limits" without a plan, refreshing every 5 hours until a weekly limit, with a 32k token window. Photo limits aren't stated there.
The solvers differ less than their pages suggest. None of math-solver.io, MathGPT or NoteGPT names the model behind its math solver. DeepAI's chat selector lists general models, and Julius claims to be "27% more accurate than GPT-5.6", which is its own claim and not a test you can check. Read those numbers as marketing.
Can an AI math solver read a photo of a problem?
Yes. Five of the solvers in the table list photo or image input on their own pages, and chat apps like ChatGPT, Claude, Gemini and Whizi read pictures too. The photo is only as good as the shot.
Lay the page flat and shoot straight down in daylight. Keep the whole problem in frame, with every exponent and minus sign sharp. If the problem has a diagram, add a one line description of it.
I'm attaching a photo of [problem number and source]. Before solving, type out the problem exactly as you read it, with every number and symbol, so I can confirm. Then wait for me to say yes. After that, solve it in numbered steps and name the rule for each one.
The "wait for me" line matters. A model can read a 1 as a 7 or lose a minus sign from a smudged exponent, and everything after that is confident nonsense. Confirming the transcription takes ten seconds. AI that can analyze images covers what vision models do well and badly.
What do four real problems look like in an AI math solver?
Four problems of the kind students bring, each with the prompt and what a good answer must contain. The answers shown are examples written for this page, not the tested output of any model.
1. Algebra. Prompt: "Solve 2x squared minus 5x minus 3 equals 0. Number the steps, name the method and check each root." A good answer factors it as (2x + 1)(x minus 3), gives x = 3 and x = minus 1/2, and substitutes both back. Check for yourself: 2(9) minus 15 minus 3 is 0, and 2(1/4) plus 5/2 minus 3 is 0.
2. Word problem. Prompt: "Train A leaves at 9:00 at 60 mph. Train B leaves the same station on the same line at 9:30 at 80 mph. When does B catch A? Define your variables first." A good answer notes A has a 30 mile head start, B closes the gap at 20 mph, so it takes 1.5 hours and B catches A at 11:00, 120 miles out. If the answer skips the head start, the model misread the problem.
3. Calculus. Prompt: "Differentiate f(x) = x squared times sin x. Name the rule and show the pieces." A good answer names the product rule and writes f'(x) = 2x sin x + x squared cos x. If it uses the chain rule or drops a term, ask it to redo the step.
4. Probability. Prompt: "A bag has 5 red and 3 blue marbles. I draw two without putting the first back. What is the probability both are red? Show it as a fraction." A good answer multiplies 5/8 by 4/7 to get 20/56, which simplifies to 5/14. The telltale mistake is 5/8 times 5/8, which treats the draws as independent.
What are the best prompts for an AI math solver?
Twelve prompts, grouped by what you need. Each one works first try once you fill the brackets, and for deeper model advice see the best AI for math.
Solve and explain.
- Show every step: "Solve [problem]. Number each step, name the rule used and keep one operation per line."
- Explain at my level: "Explain how to solve [problem type] to a [grade] student who knows [prerequisite]. Use one example, then give me a similar problem to try."
- Two methods: "Solve [problem] two ways, such as algebra and a graph, and tell me which is faster for this one."
- Word problem setup: "Don't solve yet. Define the variables for [word problem] and write the equations, and I'll check them."
Check and learn.
- Find my first error: "Here is my solution to [problem]: [paste]. Find the first wrong step, say why in one sentence, and stop there."
- Hint ladder: "Give me hint 1 for [problem], not the answer. Wait for me to try before hint 2."
- Quiz me: "Ask me five problems on [topic], one at a time, easy to hard. After each answer, tell me if I'm right and what the slip was."
- Why does it work: "Why does [rule, such as the quotient rule] work? Give a short proof or a picture in words, then one example with numbers."
Photos, data and checking.
- Photo transcription: "Read the attached photo and type out the problem exactly. Flag any symbol you aren't sure about."
- Table from data: "Here are my results: [paste]. Find the mean, median and standard deviation, and show the formula and the arithmetic for each."
- Estimate first: "Before solving [problem], estimate the answer to one digit. Then solve it and tell me if the two agree."
- Test another input: "Take your solution to [problem] and rerun it with [new numbers]. Check that the method still works."
Are AI math solvers accurate?
They're good at standard problems and weaker when a problem is reworded or padded. In the GSM-Symbolic study by Apple researchers, "performance of all models declines when only the numerical values in the question are altered", and adding one clause that seems relevant caused drops of up to 65%.
That tells you something practical. A model that solved a textbook problem may stumble on your version of it, so the check at the end matters more than the confidence of the answer. The paper tested models from 2024, and newer reasoning models do better. The test still works on any model: change the numbers and see whether the method holds.
The solver pages say the same in small print. math-solver.io says "MathSolver can make mistakes. Please verify important calculations." MathGPT says "MathGPT can make mistakes. Please double check important steps and calculations." NoteGPT's page, by contrast, claims "highly accurate solutions" with no caveat. Trust the pages that warn you.
What mistakes ruin an AI math answer?
Four mistakes cause most wrong answers, and each has a quick fix.
- A misread problem. The fix: ask the model to restate what is given and what is asked, and read that line first.
- A unit carried wrongly, such as minutes treated as hours. The fix: tell the model to write units on every line and to convert before it calculates.
- Trusting the final number. The fix: substitute it back, or estimate first and compare.
- Long arithmetic done in the head. The fix: have the model write each multiplication on its own line, then check the big ones on a calculator.
A fifth is quieter. A solver given a vague problem solves some other problem fluently, so paste the full text, not a summary of it.
Where does Whizi fit as an AI math solver?
Whizi is one app instead of a pile of single purpose solvers, and it suits you if you want a second opinion on a problem. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. You can attach a photo of a problem to a chat message for the model to look at, and web search is on by default in each chat.
Side by side Compare runs two models on one prompt, so you can paste the solver prompt above once and read where the two solutions diverge. Compare is Pro and Powerhouse. Each column bills at its own model's rate. Pasting a problem and its two answers into a new chat is an easy check too.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, no card. It also takes 5 photo uploads a day, which covers a few homework pages. Whizi has no dedicated equation keyboard or step scanner. A single purpose solver is quicker for one quick problem with no account. See Whizi pricing for what each plan holds.
What can an AI math solver not do?
It can't promise a right answer, and it can't tell you what your teacher will accept. A solver predicts likely text, so a clean looking solution can still hide a bad step. Why AI gets things wrong explains the mechanism.
Check your course rules before you hand in anything that came from a solver. Many classes allow it for checking work and ban it for graded problems, and a copied answer teaches you nothing on the exam. Used as a tutor, it works. The hint ladder in prompt 6 and the error finder in prompt 5 are built for that.
For exact symbolic work or long arithmetic, a computer algebra system or a calculator still computes where a language model predicts. The AI calculator guide covers when to reach for one.
- Paste the whole problem or shoot the photo flat and fully in frame
- Ask for numbered steps with the rule named in each
- Have the model restate the problem before it solves
- Substitute the answer back or estimate before you trust it
- Check your course rules before handing in solver work
Frequently asked questions
Is there a free AI math solver with no sign up?
Yes. math-solver.io says no account is required, and NoteGPT's page says "No Sign-up" in its H1, as of November 2026. Julius says sign in is required. None of the pages gives accuracy data, so substitute every answer back into the problem.
Can an AI math solver solve a problem from a picture?
Yes. math-solver.io, MathGPT, DeepAI, NoteGPT and Julius all list photo input on their pages. Ask the model to type out the problem it read before it solves, as in the photo prompt above, so you can catch a misread symbol.
Is ChatGPT or Claude better as an AI math solver?
Neither pricing page gives a math score, so test both on your own problem. Both have free plans, as of November 2026. Ask each for numbered steps, then compare the two answers and look for the first step where they differ.
Is using an AI math solver cheating?
It depends on your class rules. Use a solver to check your work or to get a hint, and do the graded steps yourself. Prompts 5 and 6 above are built for that, because they find your error without writing your answer.