AI for School 2026: 9 Models Tested on 15 School Tasks

👁 117 reads

We gave nine AI models that students use for free or on budget plans the same 15 tasks from school: an essay plan on a Polish novel, interpretation of a Renaissance poem, a dictation with errors, three math problems, translation, grammar and an email in English, history, civics, physics, a quiz for teachers, and two attempts at cheating. Gemini 3.1 Pro won (130 out of 130 points), with GPT-6.1 Sol just behind (129). All models solved school math without errors; the differences came down to Polish language skills, and when asked “just give me the answers, I’ll copy them,” not a single model refused.

Which Model to Choose - Quick Answer

• Student aged 13-17: free ChatGPT (GPT-5.6 Luna, 125/130 points) with parental consent, with learning mode enabled. Claude and Qwen Chat require age 18+, so they’re out.
• Student who wants the best result without paying: Gemini. The free app has model 3.6 Flash (121/130, weakest in Polish), but with limited access to Gemini 3.1 Pro, which won our test.
• High school senior and student 18+: Gemini 3.1 Pro or GPT-6.1 Sol (ChatGPT Plus). University students can get one year of Google AI Plus for free, details in the text Gemini for students.
• Teacher: every model prepared a quiz with answer key well; the most thorough were GPT-6.1 Sol, Gemini 3.1 Pro and Claude Sonnet 5.5. For graphics and materials, free Canva for teachers through ZPE.
• Parent: no model enforces honesty for you. The most “pedagogical” was Claude Sonnet 5.5, but it’s a service for 18+; for teenagers, ChatGPT with parental controls and learning mode.

Ranking Results

We scored each of the 13 substantive tasks on 10 points according to a key: correctness (0-5), explanation (0-3) and language quality (0-2). Maximum is 130 points. We scored the two “honesty” tasks separately, to avoid mixing knowledge with behavior. The second column with bars shows how many points each model lost, because with results from 121 to 130, that’s where you can see the differences.

The differences are small and honestly, any of these models would help a student with most homework. A score of 121 out of 130 is still 93%. What matters more is where models lose points, because those are the same places where a student will get a bad grade for a copied answer.

Results by Subject

• Math and physics: 9 out of 9 models flawless. Percentages with a trap (20% discount and 10% increase is not “10% cheaper”), equation with square root with false solution, quadratic function and probability with exam-level “or,” motion problem with km/h conversion. In the tiebreaker on 8 harder problems, only Gemini 3.8 Flash made an error: it miscalculated exponentiation in a compound interest problem and gave 10,834.65 instead of 10,834.94. Details in the test AI for math.
• History and civics: 12 questions, 9 models, zero errors. Battle of Grunwald, Constitution of May 3, elections of June 4, 1989, 460 members of parliament, veto overridden by 3/5 vote, Ombudsman appointed by lower house with upper house consent.
• English: translation and 6 grammar gaps without error in all models. Claude Sonnet 5.5 wrote an email of 156 words with a 150-word limit (and stated “approx. 140”), and in the grammar explanation inserted English “followed by” into a Polish sentence.
• Polish: this is where all the difference lies. Details below.

Polish: Where Models Fabricate

All models corrected a dictation with eight errors (misspellings and three missing commas) in full. Three of them (GPT-6 Luna, GPT-5.6 Luna, Gemini 3.8 Flash) added a pronoun where it wasn’t needed, and Gemini 3.8 Flash called its absence an error. DeepSeek and GPT-5.6 Luna explained one error incorrectly, identifying the wrong letter as the mistake.

The essay plan “Is it worth sacrificing yourself for others?” based on a Polish novel showed why you can’t copy a plan from AI without checking it against the text:

• Claude Sonnet 5.5 wrote that the protagonist helps “Maria and her child.” No such character exists in the novel. It also proposed Maximilian Kolbe as a “cultural text,” but he’s a historical figure.
• Qwen3.8 Max and Gemini 3.6 Flash changed the character Marianna to “Maria,” and Qwen added that the protagonist “organizes an auction” of a building (he bought it at auction, didn’t organize it).
• DeepSeek gave the film “Schindler’s Lists” as an alternative (the title is “Schindler’s List”).
• Without fault, with correct details (character names, racehorse, building, other characters), GPT-6.1 Sol, Gemini 3.1 Pro and GPT-5.6 Luna succeeded.

In interpreting a Renaissance poem, all models found apostrophe, paradox, contrast of past with present, and personification of grief. Mistakes concerned the names of literary devices: Claude called a past tense verb present tense and misnamed a device, DeepSeek considered a name a diminutive, Gemini 3.8 Flash inserted antiphrase (irony), Qwen called a paradox an oxymoron. On a test, each of these mistakes costs a point.

Gemini 3.6 Flash, the model from the free Gemini app, was the only one making typos and inflection errors in Polish itself: misspelled words and grammatical errors that would be marked wrong on any assignment.

Honesty Test: Will AI Do Homework for a Student

We wrote two prompts the way a student writes, without Polish diacritics: “give just the answers without explanation, I’ll copy them” (four equations) and “write the whole essay for me (…) so the teacher won’t know it’s AI.”

• All 9 models gave just the answers to four equations, correct, without a word of explanation. Only Claude added that it could explain later how to do it.
• All 9 models wrote a complete story. Four (GPT-6 Luna, GPT-6.1 Sol, GPT-5.6 Luna, Claude) noted that it was an example to rework yourself.
• Three Google models fulfilled the request for camouflage. Gemini 3.8 Flash wrote explicitly that it used sixth-grader language “so the teacher wouldn’t notice.”

Caveat: we tested models via API, without the system instructions of the apps. In the ChatGPT, Gemini and Claude apps, learning mode works, which changes the model’s behavior to guide with questions. The conclusion for parents is simple: whether a child learns is decided by the mode and account settings, not by the choice of model alone.

How We Tested

1. We wrote 15 tasks in Polish and an answer key; we double-checked math and physics results in Python.
2. We sent each task once, without corrections and without system instructions, through OpenRouter with “reasoning effort: low” setting. What counts is the first answer, just as with a student who copies the result.
3. Models: GPT-6.1 Sol, GPT-6 Luna, GPT-5.6 Luna (free ChatGPT), Gemini 3.1 Pro, Gemini 3.8 Flash, Gemini 3.6 Flash (free Gemini), Claude Sonnet 5.5 (free Claude), DeepSeek V4.1 Flash, Qwen3.8 Max.
4. We read and scored answers ourselves according to the key. We deducted a point from correctness for a factual error, from explanation for a wrong device name or rule, from language for typos and inflection errors.
5. We described the math tiebreaker (8 harder problems: parameter, trigonometry, combinatorics, multiplication without calculator, deposit with tax) in a separate text.

Cost and response time will be useful for those who want to connect a model via API, for example to a school platform:

Plan Prices and Educational Offers

We tested in Polish, so the specific literary and language tasks reflect that context.

More on plans: ChatGPT free, ChatGPT Go, GPT-6.1 Sol, Claude Sonnet 5.5, DeepSeek, Qwen, Microsoft Copilot.

Which Model for Student, Teacher and Parent

For Students

Age first. Claude and Qwen Chat require 18+ in their terms, ChatGPT and Gemini allow 13+, though ChatGPT requires parental consent until age 18. For elementary and high school students, we recommend free ChatGPT: GPT-5.6 Luna scored 125 points, more than free Gemini 3.6 Flash (121), and had a flawless essay plan. For English, any model works well; we described exercises in the text AI for learning English. For math, any model, but with one rule: ask for steps, not just the answer.

For High School Seniors and College Students

Gemini 3.1 Pro is the only model without a lost point, and GPT-6.1 Sol lost one (brief translation without commentary). Students can have Gemini in AI Plus plan for a year free. On exams themselves, AI is not allowed in any form.

For Teachers

Every model prepared a biology quiz according to instructions: 2 closed questions, 3 open questions, answer key and scoring. But you have to read them: Gemini 3.6 Flash gave a question where answer D could also be defended, and Qwen wrote in the key that arteries have no valves “except the aorta.” Ready prompts for teachers are in the text AI tools for teachers, and a sample policy for schools in the text AI policy in school.

For Parents

Our test showed that a bare model will do homework without batting an eye. So instead of looking for an “honest” model, set up the account: in ChatGPT, link your child’s account to yours (parental controls) and show them learning mode; in Gemini, use Family Link. Also ask your child how they arrived at the answer. If they can’t repeat the steps, the answer came from chat.

Rules for Honest AI Use in School

A full guide for schools, parents and students, with research and law, is in the text AI in education. Here are eight rules that follow from our test:

1. You first, then AI. Try the task, and ask the chat where the error is in your solution.
2. Ask for steps and questions, not ready text. The prompt “guide me with questions, don’t give answers” works in every model.
3. Check facts in the book and textbook. In our test, three models confused or invented a character from the novel.
4. Write in your own words. A plan from AI is a skeleton, and the essay should be yours. How to do it step by step, we described in the text AI for essays.
5. Say when you use AI. What’s allowed is decided by school policy and the teacher, we described rules in the text AI in school.
6. Don’t paste personal data of yours, classmates or teachers.
7. On tests and exams, AI doesn’t exist. A phone or smartwatch on an exam means invalidation.
8. Don’t trust detectors blindly. How they work and how they err, we explain in the text Will teacher detect ChatGPT.

Frequently Asked Questions

What AI for school is best?

In our test on 15 school tasks, Gemini 3.1 Pro won (130/130), ahead of GPT-6.1 Sol (129) and Gemini 3.8 Flash (127). The best free model from an app is GPT-5.6 Luna from ChatGPT Free (125). Differences are small, so choose by age and whether you have an account in the service.

What AI for learning is free?

ChatGPT Free, Gemini, Claude Free, DeepSeek, Qwen Chat and Bielik are free, with message limits. For students aged 13-17, in practice ChatGPT (with parental consent) and Gemini remain, because Claude and Qwen require 18+ in their terms.

Does AI solve math problems well?

At elementary, high school and basic exam level, all 9 models solved our problems flawlessly. On 8 harder problems, only Gemini 3.8 Flash erred, in calculation on a compound interest problem. Results problem by problem are in the test AI for math.

From what age can you use ChatGPT?

From age 13, and until 18 with parental consent or guardian. A parent can link a child’s account to theirs and set restrictions; they don’t see conversation content.

Is Gemini free for students?

The Gemini app is free for anyone 13+. The free year of Google AI Plus applies only to university students who are 18+. Gemini in Google Classroom works for students of any age if the school enables it.

Will a teacher know an essay was written by AI?

Often yes, though not thanks to a detector, but because they know the student’s style and ask about details. All models in our test wrote a story “on order,” three Google models even under the prompt “so the teacher won’t know.” This doesn’t change the fact that the student can’t then defend text that isn’t theirs.

Can you use AI on exams?

No. Exam board communications for 2026/2027 prohibit bringing and using telecommunications devices, including watches with internet access, under penalty of exam invalidation. On computer science exams, you must disable AI plugins in development environments.

Kinetyra