AI Study Modes: ChatGPT vs Gemini vs Claude (2026)

Why Feature Checklists Miss the Point

Which AI is best for studying in 2026? For retention, Claude's Learning mode is the strongest, because it withholds answers the longest and the memory research says retrieving beats being told. For a first pass at a hard concept, Gemini explains and diagrams best, and its study notebooks are the best tool for finding your weak spots before an exam. For free, everyday self-study, ChatGPT Study Mode wins: it's on every plan, remembers you across sessions, and now writes quizzes and flashcards on request.

The rest of this article shows why those answers hold, and where each one breaks.

In the space of just over four months in 2025, all three major AI labs shipped a dedicated study mode. Anthropic went first, with Learning mode inside Claude for Education in April, and opened it to every Claude user that August. OpenAI launched ChatGPT Study Mode on July 29, and Google followed with Gemini Guided Learning on August 6.

The comparisons arrived on schedule, and nearly all of them follow the same template: which mode has quizzes, which has flashcards, which integrates with your LMS, which costs what.

But a feature checklist can't answer the question a student actually has: if I study with this thing for a semester, will I know more at the end? That question has an evidence base. Cognitive psychology has spent five decades establishing what produces durable learning: retrieval practice, productive difficulty, spaced exposure, targeted feedback, and accurate self-monitoring. A tutoring product either pushes you toward those things or it doesn't. Since 2025 there have also been randomized trials of AI tutors themselves, including two that vendors ran on their own study modes.

So this comparison runs the three modes against that evidence base instead. We'll cover the features, because availability and price matter. But the spine of the article is a retention scorecard: does each mode make you retrieve, does it let you struggle, does it space anything, and how easy is it to bail out to a direct answer when the struggle gets uncomfortable?

One scope note before we start. This article compares the dedicated study modes, not the underlying models, apart from a short guide to picking a model inside each tool. If you want the model-versus-model comparison for learning tasks, we've written that separately in Claude vs ChatGPT for studying. The study modes sit on top of those models and change their behavior, sometimes dramatically, so the rankings here don't simply mirror the model rankings there.

---

The Three Modes at a Glance (October 2026)

Here's the factual layer, current as of October 2026.

Three things jump out of this table.

First, price still isn't a tiebreaker. ChatGPT Study Mode has been free everywhere since launch, and OpenAI's 2026 back-to-school offer gives students at US degree-granting institutions four free months of Plus if they claim it by October 31, 2026. Google went bigger on August 19, 2026: eligible US college students get Google AI Pro free for 12 months (normally $19.99 a month), and students in more than 140 other markets get AI Plus free for a year (normally $4.99 a month or the local equivalent), redeemable through December 31, 2026. One catch arrived in October: Google's plan changes take the Pro model out of AI Plus, so the free international year buys more usage, not the strongest Gemini. Claude's Learning mode costs nothing extra on any plan, though the full Claude for Education experience with Canvas and Wiley access still comes through your institution or not at all. Choose on pedagogy, not price.

Second, these modes are thinner than they look. In April 2026, ChatGPT's Study option vanished from the tools menu for some users without any announcement, and a Hacker News thread titled "OpenAI silently removed Study Mode from ChatGPT" climbed to 189 points. The mode hadn't been deleted, only hidden: it still opened from a typed shortcut, and OpenAI's release notes never mentioned the change. That's the nature of a behavior layer: Study Mode launched as custom system instructions on top of the base model, OpenAI said it would train the behavior into its main models once it learned what works, and it still runs with any model you can pick. Claude's version moved too: in mid-2026, Anthropic retired its styles menu and turned Learning into a skill. "Mode" still means "behavior layer," not "different brain."

Third, the three companies are aiming at different buyers. OpenAI is going wide and free, and now reaches teens through parent-set study hours. Google is bundling learning into its subscriptions, Google Classroom, and even Search, where AI Mode added free practice quizzes for tests like the SAT, ACT, and MCAT in August 2026. Anthropic is selling to provosts. That difference shapes the products more than any single feature does.

---

ChatGPT Study Mode: The Adaptive Generalist

Toggle Study Mode on and ChatGPT stops answering questions and starts asking them. Ask "how does photosynthesis work?" and instead of a four-paragraph explanation, you get a question about what you already know, then an explanation calibrated to your answer, then a check: can you explain the light-dependent reactions back?

To turn it on, type @study in the message box, or tap + and pick Study. On the web you can also go straight to chatgpt.com/studymode. It works with any model you have access to, but not inside Projects, custom GPTs, or temporary chats, so keep your tutoring in a normal chat.

What changed in 2026 is the toolkit around the mode. On August 14, OpenAI added interactive quizzes you answer directly in the conversation. On September 22, it added flashcards: ask for a deck or upload your notes, flip each card, mark it as known or as one to practice again, and the deck saves to a library you can reopen. Both are app-wide features rather than parts of Study Mode: flashcards work on every plan, and quizzes on every consumer plan and Edu. ChatGPT's memory means it can remember that you confused mitosis and meiosis last week and quiz you on it again, if you come back and ask. And on October 7, OpenAI began rolling out GPT-6 in chat: Plus, Pro, Business, and Enterprise get GPT-6 Sol, and Free and Go users move to GPT-6 Luna from October 8, though the tutoring behavior matters more here than the raw model.

Scale is the other story. OpenAI says more than 180,000 teens used Study Mode in a single week, and parents can set study hours that start a teen's new chats in Study Mode. In some eligible ChatGPT Edu workspaces, administrators can go further and require the mode in regular chats, though not inside GPTs or Projects. OpenAI is also measuring whether any of this works: with Stanford's SCALE Initiative and the University of Tartu it built a learning-outcomes measurement suite, and its own early randomized results appear in the evidence section below.

The strengths are reach and adaptivity. It's free on every plan, works in every region, and the personalization is the best of the three because ChatGPT's memory system is the most developed. For a self-directed learner grinding through a topic over weeks, the continuity is genuinely useful.

The weakness is the same thing that let it slip out of the menu unannounced: it's a behavior layer, and the behavior bends under pressure. Push back ("just tell me the answer, I'm in a hurry") and Study Mode can fold; OpenAI's own help page says it may sometimes give a direct answer. The full answer-giving ChatGPT is a click away, in the same window, with no friction. We'll see why that matters in the scorecard.

---

Gemini Guided Learning: The Visual Scaffolder

Gemini's take is less conversational and more structural. Ask Guided Learning about a topic and it builds you a scaffold: the concept broken into ordered chunks, diagrams and images generated inline, relevant YouTube videos embedded where they help, and a quiz at the end of a sequence. It feels less like a tutor asking you questions and more like a textbook chapter assembling itself around your question.

To open it on the web, click + (Add files) in the prompt box, choose More tools, then Guided Learning. It's free with a Google sign-in, and Google is still rolling it out to the mobile apps.

The textbook feel isn't a criticism. For visual and structural learners, and for subjects where a diagram does the work of five paragraphs (organic chemistry mechanisms, supply and demand curves, anatomy), Gemini's presentation layer is the best of the three. The embedded video selection is a real advantage no competitor matches, because Google owns YouTube.

The bigger 2026 change sits next to Guided Learning rather than inside it. Study notebooks, launched on the web on June 25, 2026 and on mobile on August 19, start from your material: type a topic or upload a syllabus or notes, and Gemini offers a diagnostic quiz, then builds a plan of short interactive lessons around your weak spots, each with its own practice quiz. A progress view sorts every objective into Strengths, Focus areas, or Not started and updates as you finish quizzes. The notebooks also sync with Gemini Notebook (the product formerly called NotebookLM), where you can turn the same material into flashcards. Study notebooks are the closest any of these products comes to shaping a whole course rather than a single session, and they lean on testing rather than telling.

Gemini has an LMS story now, too. Since August 10, 2026, Gemini in Google Classroom has been open to K-12 and higher-education students of all ages whose school admins allow it, so a student can build quizzes and flashcards from course materials or open Guided Learning without leaving Classroom.

Underneath, the plumbing keeps moving. On October 9, 2026, Google began switching free users to Auto, which sends most prompts to the fast Flash-Lite model and escalates harder ones to Flash or Pro. AI Pro subscribers can pick 3.8 Flash or 3.1 Pro directly. Google has also put Guided Learning through a randomized trial in Sierra Leone schools, covered in the evidence section below.

The pedagogical weakness of Guided Learning itself hasn't changed. It presents more than it asks. The quizzes exist, but they arrive at the end of a well-organized explanation, which means you've been reading, not retrieving. Reading a beautifully scaffolded explanation feels like learning. The research covered below says that feeling is exactly the thing to distrust. Study notebooks fix much of this, which is why the scorecard treats them as part of Gemini's study toolkit.

---

Claude Learning Mode: The Socratic Purist

Claude's Learning mode is the most opinionated of the three. Ask it how to solve a calculus problem and it will not solve the problem. It asks what you've tried, where you got stuck, and what you think the next step might be. It keeps asking. Of the three modes, it resists the "just give me the answer" pressure the longest.

How you switch it on has changed twice. Learning mode began in Claude for Education in April 2025 as a mode inside Projects, then reached every user in August 2025 as a Learning option in Claude's style menu. In mid-2026, Anthropic retired styles and made Learning a skill. Add it once under Customize, then Skills (search Browse skills for "Learning"), and turn it on in any conversation. Skills are available on every plan, Free included, as long as code execution is enabled in settings. Developers get the same idea in Claude Code, whose Learning output style leaves small, meaningful pieces of code marked TODO(human) for you to write yourself.

Anthropic's education push is aimed at institutions, and it shows. Claude for Education includes Canvas LTI integration, so students work with Claude inside their actual courses without switching apps. Connectors to Wiley put peer-reviewed reference content in the conversation, and Panopto integration brings in lecture transcripts. Student conversations are excluded from model training by default, which is the kind of clause university counsel actually reads.

The institutional bet keeps paying. Northeastern University runs Claude for students and staff across its 13 global campuses, which serve nearly 49,000 students. LSE and Champlain College signed on at launch in April 2025, Northumbria followed that July, and Syracuse went campus-wide in September 2025. The University of Pittsburgh (October 2025) and Dartmouth (December 2025) came in through Amazon Web Services. In January 2026, Anthropic announced a partnership with Teach For All that opens AI training to its network of more than 100,000 teachers and alumni in 63 countries, a signal that the education strategy extends well beyond wealthy universities. In summer 2026, Anthropic added Claude for Teachers, a free year of Claude for verified US K-12 teachers who sign up by June 30, 2027.

Under the hood, the lineup turned over between September and October 2026: Fable 5.1 arrived on September 1, Opus 5.5 on September 22, Sonnet 5.5 on September 28, and Haiku 5.5 on October 7. Free users get Sonnet and Haiku, and Pro adds Opus 5.5, which Anthropic says performs at the level of Fable 5.1 on most work.

The strengths follow from the philosophy: the strictest retrieval-forcing behavior of the three, and a ready-made Canvas integration for campuses. The weaknesses follow too. Without a campus deal, you get the Learning skill but none of the integrations. There's still no learning-progress dashboard: memory now works on every plan, and a monthly recap (in beta) shows which topics you spent time on, but nothing tracks what you've mastered. And the strictness cuts both ways: students who just need a quick factual answer at 11 p.m. find the relentless questioning genuinely annoying, and the annoyed ones switch it off.

---

Which Model Is Best for Studying in Each Tool

The best model for studying is GPT-6 Sol in ChatGPT, Claude Opus 5.5 in Claude, and Gemini 3.1 Pro in Gemini, though the free default in each is enough for most coursework. A study mode is a behavior layer, and the model underneath sets the ceiling on how well the tutor reasons, explains, and catches its own mistakes. All three lineups changed between September and October 2026, so here's where each one stands.

The mode's questioning discipline matters more than the tier. GPT-6 Luna on the free plan tutors well for typical homework, Claude's Sonnet handles the majority of coursework, and Gemini's Auto setting escalates hard prompts on its own. Reach for the stronger models when the reasoning gets hard: multi-step math and physics, graduate-level proofs, or feedback on a document longer than about ten pages, where the stronger model's ability to hold context and self-correct starts to show. Students rarely need the tiers above those (GPT-6 Pro on ChatGPT's Pro plans from $100 a month, Claude Fable 5.1 on Max, Google AI Ultra).

---

What Learning Science and the New Trials Say a Tutor Must Do

Now the evidence base, because this is where the three modes stop looking interchangeable.

Retrieval practice is the single best-documented finding in learning science. Roediger and Karpicke's 2006 experiments in Psychological Science had students either restudy a passage or study it once and then test themselves on it. Five minutes later, the restudiers looked better, 81% to 75%. A week later the result flipped hard: the students who had practiced retrieval recalled 56% against the restudiers' 42%. The act of pulling information out of memory strengthens it in a way that re-reading does not. We cover the practical method in our guide to active recall. For a study mode, the implication is blunt: every time the AI explains something you could have retrieved, it took the rep for you.

Desirable difficulties are the second pillar. Robert Bjork's research program at UCLA established that conditions which make learning feel harder and slower (testing before you're ready, spacing sessions apart, interleaving topics) produce better long-term retention than conditions that feel smooth. Fluent, frictionless study is a reliable way to feel prepared and not be. We've written a full piece on desirable difficulties, but the one-line version for this comparison: a tutor that removes all struggle is removing the learning along with the discomfort.

Cognitive offloading is the failure mode. Michael Gerlich's 2025 study in Societies, a survey of 666 participants in the UK, found that heavier AI tool use correlated with lower critical-thinking scores, partly through cognitive offloading, and participants aged 17 to 25 reported the most AI use and scored lowest. It's a correlation, not proof of cause. The MIT Media Lab's "Your Brain on ChatGPT" preprint (Kosmyna and colleagues, 2025) put EEG caps on 54 essay writers and found the ChatGPT-assisted group showed the weakest neural connectivity. In the first session, 83% of that group couldn't correctly quote a sentence from the essay they had written minutes earlier, against 11% in the other groups. The paper hasn't been peer reviewed, and its methods have drawn criticism, but the researchers' name for the pattern stuck: "cognitive debt," borrowed performance now, paid for in capability later. We go deeper on this in the AI thinking trap.

The trials say guardrails decide whether AI costs you learning. Until 2025, the case for study modes rested on lab findings like these. Then randomized trials of AI tutors started landing, and in the one that pitted answers against hints, the results split along exactly the line the study modes draw.

• Answers versus hints: In a Turkish high school, Hamsa Bastani and colleagues gave nearly 1,000 students plain GPT-4, a tutor version, or no AI during math practice, and published the results in PNAS in June 2025. With AI available, plain GPT-4 raised practice scores 48% and the tutor 127%. Then the AI was taken away for the exam. The plain GPT-4 group scored 17% worse than students who never had AI, while the tutor group was statistically indistinguishable from them. The tutor was prompted to give hints without handing over answers, and teachers fed it worked solutions and common mistakes.
• A tutor built to pace the answer: At Harvard, Gregory Kestin's team built a physics tutor that gave away one step at a time, never the full solution in a single message, and handed over the answer only when a student demanded it. In a 194-student trial published in Scientific Reports in June 2025, median learning gains with the AI tutor were more than double those from an in-class active-learning lesson, in a median 49 minutes against a 60-minute class.
• In the field: A World Bank team ran a six-week after-school English program in Nigeria in which students worked with Microsoft Copilot, powered by GPT-4, under teacher supervision. Scores on a combined test of English, AI knowledge, and digital skills rose 0.31 standard deviations; English alone, the main outcome, rose 0.23.
• The vendors' own modes: Google's LearnLM team ran a randomized trial of Guided Learning across 48 math classes in 12 schools in Sierra Leone. Half the classes, 871 of the 1,763 students, used it in lessons for about two months and outscored the other half by 0.26 standard deviations, which the report equates to more than a year of typical learning. It's a technical report, not yet peer reviewed. OpenAI reported early results from its own randomized study of more than 300 college students preparing for exams: students given access to Study Mode scored roughly 15% higher than a no-AI control group in microeconomics, while in neuroscience the difference wasn't statistically distinguishable, and OpenAI says the analysis is still underway. Anthropic hasn't published an outcome study of Learning mode.

We walk through the Harvard, Turkish, and Nigerian trials in detail in our guide to whether AI tutors actually work. For this comparison, one line is enough: the same model harmed learning when it handed over answers and did no measurable harm when it gave hints instead.

Put those lines of research together and you get a concrete checklist for evaluating an AI tutor:

1. Retrieval prompts: Does it make you produce answers from memory, or does it present information for you to recognize?
2. Spacing: Does anything bring material back days later, when forgetting has set in and retrieval is hard enough to count?
3. Feedback quality: When you're wrong, does it diagnose why you're wrong, or just supply the correction?
4. Metacognition support: Does it help you calibrate what you actually know versus what merely feels familiar?
5. Escape-hatch risk: How easily can you bail out of the productive struggle into a direct answer?

That fifth item deserves emphasis because no checklist comparison mentions it, and it's the closest thing to the variable the Bastani trial isolated: answers on tap versus hints only. The moment of maximum frustration in studying is also the moment of maximum learning, if you stay in it. Every one of these products puts a "just answer me" mode within one click of that moment.

---

The Retention Scorecard

Scoring the three against the five criteria, based on how each behaves as of October 2026. Each one is scored on the study toolkit a student actually uses: ChatGPT's Study Mode with the app's quizzes and flashcards, Gemini's Guided Learning with study notebooks, and Claude's Learning mode. Three points per criterion, 15 possible.

Read the totals loosely. No trial has compared these three products head to head, so the scores are our judgment against the research, not a measured result, and the right mode still depends on the task. But three findings survive any reasonable re-scoring.

Claude wins on pedagogy because its default behavior is the closest match to what the testing-effect literature prescribes, and to the hints-not-answers design that protected students in the Bastani trial. The thing students complain about, the refusal to just answer, is the active ingredient. That's the entire design thesis of Learning mode, and the research backs it.

Gemini owes its tie with ChatGPT to study notebooks, not Guided Learning. Guided Learning on its own is still the best explainer of the three, and explanation is the part of tutoring that learning science values least: a gorgeous scaffold you read is still re-reading. A notebook that offers to test you first, labels what you don't know, and keeps that map between sessions is a different and better design.

Nobody does spacing well. This is still the open gap in the entire category. The forgetting curve does most of its damage between sessions, and durable learning depends on re-retrieving material spaced out over days and weeks. In 2026 all three products gained ways to bring material back, from flashcard decks to scheduled tasks, but every one of them waits for you to ask or to build the schedule yourself. Nothing in any of these products will, on its own, notice that Tuesday's confusion is due for review on Friday. Which leads to the next section.

---

The Half of the System No Mode Provides

Here's the framing that the checklist comparisons miss entirely: an AI study mode covers, at best, half of a learning system.

The half it covers is explanation and questioning in the moment. The half it doesn't cover is what happens to that material afterward: capturing what you learned in a durable form, encoding it through your own words, and reviewing it on a schedule that respects the forgetting curve. Flashcard decks and study notebooks help, but they live inside one vendor's app, and most study sessions still evaporate into a chat log you will never scroll back through. Ask yourself honestly when you last re-read an old AI conversation.

The fix isn't another AI feature. It's a capture habit that you own, sitting alongside whichever mode you use.

Concretely, the loop looks like this. When a study session surfaces something worth keeping (a definition that finally clicked, a worked example, a distinction you kept getting wrong), don't leave it in the chat. Save the source material with Glasp's web highlighter when you're working from articles or documentation, so the original passage and your reaction to it persist. When your course material is lecture video, YouTube Summary turns the lecture into notes you can highlight and keep instead of a 90-minute blob you'll never rewatch. If your textbooks live on Kindle, Glasp's Kindle import brings those underlines into the same library.

Then the review half: a few days later, instead of re-reading, interrogate. Glasp's AI chat works over your own highlights, so you can ask "quiz me on what I saved about enzyme kinetics this month" and get retrieval practice generated from material you personally selected and encoded. Put those sessions on your own calendar, spaced days apart, and you have the spacing mechanism the study modes lack, built on your highlights rather than a vendor's chat history. That matters more than it sounds: features in this category move without warning, as some ChatGPT users found in April when the Study option disappeared from their menu.

The division of labor is clean. Study mode for the live session: explanation, Socratic questioning, working through problems. Your own capture and review system for everything the session needs to become: encoded, spaced, retrievable. Students who run both halves get the benefit of the tutor without the cognitive debt, because the encoding work stays theirs.

---

Verdict: Which Mode for Which Job

No single winner. The answer is a mapping from task to tool.

A few caveats to go with the table.

If your university runs Claude for Education, that decision is mostly made for you, and it's a good default. The integration advantages are real, and a course built around Learning mode gives you a reason not to flip it off at midnight.

If you're price-sensitive, 2026 is unusually kind. ChatGPT Study Mode on the free plan is a remarkable amount of tutoring for nothing, and the student offers covered above make the paid tiers free for a while too. Price has stopped separating these tools, so decide on how each one teaches instead.

One thing to take from the scorecard: the mode you pick matters less than whether you stay in it when it gets uncomfortable, and whether you capture and review what it teaches you. A mediocre mode used with discipline beats the best mode with the escape hatch worn smooth.

---

Frequently Asked Questions

Is ChatGPT Study Mode good for exam prep?

Good, with one discipline requirement and one supplement. The discipline: don't switch out to normal chat when the questioning gets frustrating, because that frustration is the retrieval practice doing its work. The supplement: Study Mode won't resurface material on a spaced schedule by itself. Ask for a flashcard deck at the end of each session, then set a scheduled task (free plans allow up to three) that prompts a quiz on that material a few days later. Used that way, it's a strong and free exam prep tool.

How do I turn on ChatGPT Study Mode?

On the web, type @study in the message box and pick Study from the suggestions, or go straight to chatgpt.com/studymode. In the iOS app, tap + and choose Study; on Android, tap + and search the menu for it. It works with every model on every plan, free included, but not inside Projects, custom GPTs, or temporary chats. If the option seems to have vanished from your menu, as it did for some users in April 2026, the @study shortcut and the direct link still work.

Does Claude have a study mode?

Yes. Claude's equivalent is Learning mode, which makes Claude ask guiding questions and withhold finished answers. It started in Claude for Education in April 2025 and reached every user in August 2025. Since mid-2026 it's a skill rather than a style: open Customize, then Skills, add Learning from Browse skills, and switch it on in a conversation. Skills work on every plan, including Free, with code execution enabled. Developers can get similar behavior in Claude Code by choosing the Learning output style. Claude is the better pick for school when you want a tutor that makes you do the thinking: Learning mode scored 11 of 15 on our retention scorecard to ChatGPT Study Mode's 9. For everything else, our Claude vs ChatGPT for studying guide compares the two on eight study tasks.

Is using AI study modes cheating?

Using a study mode as designed is closer to hiring a tutor than to cheating: the mode asks you questions and makes you do the work, which is what an honest tutor does. Using the same product with the mode switched off to generate your assignment answers is a different activity with a different name. Most universities now draw the line at exactly that distinction, and some ChatGPT Edu workspaces let administrators require Study Mode in regular chats. When in doubt, your course's AI policy beats any general answer; our guide on whether using AI to study is cheating covers where the 2026 rules sit.

Which AI study mode is completely free in 2026?

All three, with caveats, and ChatGPT's is the best pick if you never plan to pay. Study Mode is free on every plan, worldwide, and works with whatever model your plan includes. Gemini Guided Learning is free with a Google sign-in, within usage limits that refresh every five hours up to a weekly cap. Claude's Learning skill works on the Free plan, though the full Claude for Education experience (Canvas, Wiley, Panopto) requires an institutional deal. On top of that, eligible college students can claim a free year of Google AI Pro in the US or AI Plus in more than 140 other markets until December 31, 2026, and students at US institutions can claim four free months of ChatGPT Plus until October 31, 2026.

Gemini Guided Learning vs ChatGPT Study Mode: which mode is better?

ChatGPT Study Mode is the stronger all-rounder for self-study: it's free for everyone, its memory, quizzes, and flashcards compound across weeks, and its Socratic questioning forces retrieval. Gemini wins when you're meeting a hard concept for the first time, because Guided Learning's diagrams and embedded videos explain better than anything else, and its study notebooks are the best tool in the category for diagnosing your gaps. For raw retention, though, neither beats Claude's Learning mode, which resists handing over answers the longest. Our Gemini vs ChatGPT for studying guide compares the two head to head, including the student offers.

What's the difference between Gemini Deep Research and Guided Learning?

Deep Research and Guided Learning point in opposite directions. Deep Research does the work for you: it searches the web (plus your Gmail, Drive, or uploaded files if you add them) and writes a long report with sources on a topic. Guided Learning makes you do the work: it breaks a subject into steps and walks you through them with questions, diagrams, and quizzes. Use Deep Research to survey sources for a paper, and Guided Learning or a study notebook when the goal is to understand and remember the material yourself.

Which ChatGPT, Gemini, or Claude model is best for studying?

The best models for studying are GPT-6 Sol on ChatGPT Plus, Gemini 3.1 Pro on AI Pro, and Claude Opus 5.5 on Claude Pro, but for most studying the free default is enough, because the study mode's behavior matters more than the raw model. As of October 2026 that means GPT-6 Luna on ChatGPT's free plan, Gemini's Auto setting on the free tier, and Claude's Sonnet on the Free plan. Step up to the paid models when the reasoning gets hard: multi-step proofs, dense STEM problem sets, or feedback on writing longer than about ten pages.

Do AI study modes actually improve learning?

Early trials say yes for AI tutors built to guide rather than just answer, or used under a teacher's direction, so the case is now experimental rather than only theoretical. In a Harvard physics trial, a step-by-step AI tutor more than doubled median learning gains compared with an active-learning class. In a Turkish high school trial, a hints-only GPT-4 tutor protected exam scores while plain GPT-4 cut them 17%. A World Bank program in Nigeria and Google's Guided Learning trial in Sierra Leone each found gains of roughly a quarter to a third of a standard deviation. The failure mode is just as well documented: heavy everyday AI use correlates with weaker critical thinking (Gerlich 2025). Study modes are an attempt to ship the designs that worked in the trials, and whether the commercial versions match the research prototypes is still being measured.

Can I just use a system prompt instead of a study mode?

Mostly, yes. ChatGPT's Study Mode launched as custom system instructions on top of the base model, and you can instruct any frontier model to tutor Socratically, refuse direct answers, and quiz you. The Harvard and Turkish trial tutors ran on GPT-4 with written prompts rather than a specially trained model, and they went further: both were given worked solutions, so they could check a student's steps instead of improvising. Two things the DIY version lacks: persistence (the modes, with their memory, saved flashcards, and notebooks, carry across sessions) and commitment (a prompt you wrote is even easier to abandon than a toggle). If you have the discipline, a custom prompt on any model gets you most of the pedagogy. The modes exist because most people, under deadline pressure, don't have the discipline.

---

Conclusion

Strip away the marketing and the three study modes are three answers to one question: how much friction should stand between a student and an answer? Claude says a lot, and the retention research agrees. Gemini says very little inside Guided Learning, then makes up for it with study notebooks that offer to test you first. ChatGPT splits the difference and gives it away free, with the best memory of the three.

The feature lists will keep getting longer, and the modes will keep converging. What won't change is the underlying science, now backed by randomized trials: you remember what you retrieve, AI that hands over answers can cost you the exam while tutors built to guide have produced real gains, and no tutor, human or AI, can do the encoding for you. Pick the mode that fits your task from the verdict table. Then build the half of the system no mode ships: capture what you learn, in your own words, somewhere you own.

That second half is what Glasp is for. Highlight what your study sessions teach you, pull your lecture videos into notes with YouTube Summary, and when the forgetting curve comes for the material next week, use Glasp's AI chat to quiz yourself on your own highlights instead of re-reading them. The AI tutor handles the conversation. Make sure you're the one who keeps the knowledge.