Statistics AI: Free Tools, 12 Prompts, 4 Worked Tests

In brief

statisticsai.io gives 20 AI answers a day on its free plan, and paid is $6 a month for 400, as of November 2026. ChatGPT and Claude have free plans too. Any can slip: in the StatQA study by Zhu and colleagues, the best GPT-4o score was 64.83%, so recompute every figure.

What prompt makes statistics AI show its work?

Paste this into Claude, ChatGPT, Gemini or Whizi and fill the brackets. It makes the model check assumptions, show the formula and report an interval, not just a verdict.

A row of five smooth black stones stepping up in height on a black plinth beside a small rainbow glass bell curve and an open notebook of blank graph paper, with a terracotta Claude starburst, a silver OpenAI knot and a rainbow glass Gemini star on their own plinths in a charcoal studio under white rim light.

Here is my data: [paste the values, or say the file is attached]. Question: [what you want to know, such as whether two groups differ]. Use [test name or "choose the right test and say why"] at alpha [0.05]. First list the assumptions and say whether my data meets them. Then show the formula, the numbers substituted in, the test statistic, the degrees of freedom, the p value and a [95]% confidence interval. End with one plain sentence on what the result means, and flag any step that is an estimate.

Each part closes a gap. Listing assumptions first catches a test that doesn't fit before it produces a number. Showing the substitution lets you spot a wrong value in ten seconds. The last line stops a vague "significant" with no size attached.

Which statistics AI is free?

Most of the tools Google ranks for this query have a free tier, and ChatGPT and Claude do too. The table lists what each tool says on its own page, read on 22 November 2026. A cell that says not stated means the page gave no figure.

ToolFree useSign upPaid priceSteps, photos, filesBest at
statisticsai.io20 AI answers a dayNot stated$6 a month for 400 answers; $4.5 a month on yearly for 4800Steps; image upload; files not statedQuick textbook problems
TutorBin statistics solverFree per its FAQ, no number givenNot stated; Sign In shownNo subscription fees statedSteps; image and camera; files not statedA free first pass with explanations
Stats AI (iPhone and iPad app)Free to download, with in app purchasesNot statedPro $4.99 weekly or $29.99 yearlySteps; photo; chat tutor; files not statedSnapping a problem from a textbook
ChatGPTFree, limited messages with uploads; data analysis marked limitedNot statedNot stated on the page I readSteps when asked; file upload limited on FreeExplaining a method in conversation
ClaudeFree, limited usage every 5 hours and every weekNot statedPro $17 a month billed annually, $20 monthlySteps when asked; page lists creating files and running code on FreeChecking a calculation with code
GeminiStandard limits, refreshing every 5 hours until a weekly limitNot statedNot statedSteps when asked; file limits not statedLong inputs within a 32k token window
Whizi10 messages a day on Whizi AI, Auto and Claude Haiku, no card; 5 photos a dayAccount neededPro $29.99 a monthSteps when asked; attach a photo, CSV or Excel file; make an Excel file in chatRunning one question past two models

Whizi loses in two places. Neither TutorBin nor statisticsai.io states that an account is needed, while Whizi wants one and a prompt like the one above. Claude's pricing page also lists running code on the free plan, a claim this page does not make for Whizi, so for a big dataset recompute the key numbers in a spreadsheet or R.

The three dedicated solvers claim a lot about accuracy and say little about limits. statisticsai.io says its solver "achieves 98% accuracy across all types of statistical problems", and TutorBin's page links a "100% Correct Solutions" feature. Neither shows a test you can check, so read those lines as marketing. None of the three states CSV upload.

Can statistics AI analyze a CSV or Excel file?

A chat app can, and the dedicated solvers mostly don't say. In Whizi you attach a CSV or an Excel .xlsx file to a message, and the model reads the values as text. The composer also takes older .xls files, but it asks you to re save them as .xlsx first.

Small files work well. A few hundred rows fit in a message, and the model can describe the columns, flag missing cells and suggest a test. For a large file, ask it to profile the data first and run the real calculation in a spreadsheet, because reading thousands of numbers as text is where a model drops one.

I've attached [file name]. Before any analysis, tell me the number of rows, the columns and their types, the count of blank cells per column and any value that looks like a typo or outlier. Then wait for me to confirm which columns to compare.

You can also go the other way. Whizi can make an Excel file in a chat, with no plan gate on it, so ask for a sheet that holds your data with the formulas written in cells, such as AVERAGE and STDEV.S. Then the arithmetic is done by Excel and you can click into every cell. The Excel AI guide covers that route.

What do four real statistics problems look like?

Four problems of the kind students and analysts bring, each with the prompt and the numbers a good answer must reach. I computed every figure with a script before writing it. The answers shown are examples written for this page, not the tested output of any model.

1. Descriptive statistics. Prompt: "Find the mean, median, sample standard deviation and a 95% confidence interval for 12, 15, 11, 14, 18, 16, 13, 15." A good answer gets a mean of 14.25, a median of 14.5 and a sample standard deviation of 2.25. The standard error is 0.796, the t value for 7 degrees of freedom is 2.365, and the interval runs from 12.37 to 16.13. The classic slip is the population standard deviation, 2.11, which divides by n instead of n minus 1.

2. Two sample t test. Prompt: "Group A scored 78, 82, 85, 91, 74, 88, 80, 79 and group B scored 84, 90, 87, 95, 83, 92, 89, 86. Run a two sided Welch t test at alpha 0.05." A good answer reports means of 82.125 and 88.25, standard deviations of 5.59 and 4.06, t of minus 2.51 on about 12.8 degrees of freedom and p of 0.027, so the groups differ at 0.05. A pooled test gives the same t with 14 degrees of freedom and p of 0.025. Both agree here, but the answer should say which one it ran.

3. Proportion z test. Prompt: "In 1,000 voters 540 back a measure. Test whether support differs from 50% at alpha 0.05." A good answer uses the standard error under the null, the square root of 0.25 divided by 1,000, which is 0.0158. That gives z of 2.53 and a two sided p of 0.011, so reject. If the answer uses 0.54 times 0.46 inside the root, the z comes out slightly different, and it is the wrong form for a test of 50%.

4. Chi square test. Prompt: "A 2 by 2 table has 30 and 20 in the first row and 20 and 30 in the second. Test independence at 0.05 and say whether you applied a continuity correction." Without correction, chi square is 4.00 with 1 degree of freedom and p of 0.0455, which rejects. With the Yates correction, it is 3.24 and p of 0.072, which does not. A good answer names the choice, because here it flips the verdict.

What are the best prompts for statistics AI?

Twelve prompts, grouped by what you need. Each works first try once you fill the brackets. For other tools in this space see the best AI for data analysis.

Choose and run a test.

  1. Pick the test: "I have [describe the data: groups, sample sizes, measured or counted] and I want to know [question]. Which test fits, and what does it assume? Don't calculate yet."
  2. Full test report: "Run a [test name] on [data]. Give hypotheses, assumptions, the statistic, degrees of freedom, p value, a confidence interval and an effect size, then one sentence on what it means."
  3. Sample size: "How many people do I need to detect a [difference] with a standard deviation of [value], power [0.8] and alpha [0.05]? Show the formula."
  4. Non normal data: "My data is [skewed or small, describe it]. Should I use a t test or a rank based test, and why? Then run the one you pick."

Understand and check.

  1. Explain a result: "Here is my output: [paste]. Explain each number to someone who has not taken statistics, and say what I can and cannot conclude."
  2. Find my error: "Here is my working: [paste]. Find the first wrong step, say why in one sentence, and stop there."
  3. Hint ladder: "Give me hint 1 for [problem], not the answer. Wait for me to try before hint 2."
  4. P value in plain words: "Explain what a p value of [value] means in my study of [topic], and list three things it does not mean."

Data, probability and study design.

  1. Clean first: "Read the attached file and list the columns, types, blank cells and outliers before analysis."
  2. Probability setup: "Don't solve yet. Define the events and the distribution for [problem], and write the formula I need."
  3. Regression read out: "Fit a line to this data: [paste x and y]. Give the slope, intercept, R squared and p value for the slope, and tell me if a straight line is a fair model."
  4. Study design: "I want to test whether [intervention] changes [outcome]. Suggest a design, name the biggest source of bias and say what I should record before I start."

How good is AI at statistics?

It is good at routine formulas and uneven at choosing the right test. The StatQA benchmark, built by Yizhang Zhu, Shiyin Du, Boyan Li, Yuyu Luo and Nan Tang, holds 11,623 statistics examples. Its authors report that "even state-of-the-art models such as GPT-4o achieve a best performance of only 64.83%". Its authors also found that LLMs "primarily make applicability errors", meaning they apply a method that does not fit the data, while humans mostly confuse which statistical task they face.

That is why the prompt above asks for assumptions before arithmetic. The paper, "Are Large Language Models Good Statisticians?", tested 2024 models, and newer ones do better, but the habit still pays: a clean looking t test on data that breaks the assumptions is a wrong answer with correct arithmetic. Why AI gets things wrong explains the mechanism.

The vendor pages give no statistics score for any of the tools in the table. Treat any "98% accurate" line as a claim and test the tool on a problem whose answer you already know, such as the four above.

What mistakes ruin a statistics AI answer?

Four mistakes cause most wrong answers, and each has a quick fix.

  1. The wrong test for the data, such as a t test on paired measurements. The fix: ask which test fits and why before you ask for numbers, and say whether the groups are paired.
  2. A summary pasted instead of the data. The fix: paste every value, so the model computes from real numbers and does not invent some.
  3. Sample and population mixed up. The fix: state "sample standard deviation" or "population", since the two differ, 2.25 against 2.11 in problem 1.
  4. A p value read as proof. The fix: ask for the effect size and the interval beside it, and for what the result does not show.

A fifth is quieter. A model asked for "the answer" may round early, so ask it to carry full precision and round only the final line.

Where does Whizi fit as a statistics AI?

Whizi is one app instead of a pile of single purpose solvers, and it suits you if you want a second opinion on a calculation. Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. You can attach a CSV, an Excel file or a photo of a problem to a chat message, 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 test prompt above once and read where the two reports differ. Compare is Pro and Powerhouse. Each column bills at its own model's rate. Pasting a result and its two answers into a new chat is another easy check.

The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, no card. Whizi has no dedicated stats keyboard or distribution calculator, and a single purpose solver is quicker for one textbook problem. See Whizi pricing for what each plan holds.

What can statistics AI not do?

It can't see how your data was collected, and it can't promise a right answer. A model predicts likely text, so a tidy solution can sit on a wrong test. It also can't tell whether your sample was biased, and no p value fixes that.

Check your course rules before you hand in anything that came from a tool. Many classes allow AI to check work and ban it for graded problems. Used as a tutor, it works. The hint ladder in prompt 7 and the error finder in prompt 6 are built for that.

For long arithmetic or a big dataset, a spreadsheet, R or a calculator still computes where a language model predicts. The AI calculator guide covers when to reach for one, and the AI math solver guide covers the algebra underneath.

Workflow checklist
  • Paste every data value or attach the file, never a summary
  • Name the test, the alpha and whether groups are paired
  • Ask for assumptions before the result
  • Recompute the mean or the test statistic yourself
  • Check your course rules before handing in AI work
Common questions

Frequently asked questions

Is there a free statistics AI?

Yes. statisticsai.io lists 20 AI answers a day on its free plan, and TutorBin's FAQ says students can use its solver for free, as of November 2026. ChatGPT and Claude have free plans too. None states a statistics accuracy score, so recompute one number.

Can AI do a t test or a chi square test?

Yes, if you paste the data and name the test. The four worked problems above give the numbers a correct answer must reach, such as t of minus 2.51 and p of 0.027 for problem 2. Ask for assumptions first, since choosing the wrong test is the usual error.

Can statistics AI read an Excel or CSV file?

Chat apps can, and the three dedicated solvers do not state it. In Whizi you attach a CSV or .xlsx file to a message. Ask for a profile of the columns and blanks first, and use a spreadsheet for the calculation on large files.

Is using statistics AI cheating?

It depends on your class rules. Use it to check your work or to get a hint, and do the graded steps yourself. Prompts 6 and 7 above find your error or give a hint without writing your answer.