Welcome back to Glasp Deep Dive. Today, three AI tutors walk into a comparison: ChatGPT study mode, Gemini Guided Learning, and Claude Learning Mode. And we're not counting features.

Because every comparison out there already does that. Quizzes, flashcards, integrations, price tiers. Useful, but it dodges the question a student actually has.

Which is brutal and simple. If I study with this thing for a semester, do I actually know more at the end? And that question has an evidence base.

Five decades of it. Cognitive psychology keeps finding the same drivers of durable learning: retrieval practice, productive difficulty, spaced exposure, targeted feedback, honest self-monitoring.

Plus one criterion almost nobody covers, the spine of this episode, the escape hatch. Every product here puts a "just answer me" mode one toggle away, and I know exactly who would use it. Me.

Hold that confession. Lay of the land first. OpenAI shipped ChatGPT study mode in July twenty twenty-five. Google followed with Gemini Guided Learning that August.

And Anthropic was actually earliest, right? Learning mode arrived inside Claude for Education back in April twenty twenty-five. Then a year of signing up universities.

Three labs, one year, three different buyers. OpenAI is going wide and free. Google is bundling learning into its consumer subscription. Anthropic is selling to provosts.

Selling to provosts is a very specific strategy, and you can feel it in the product. Before we go mode by mode though, tell the April story.

April twenty twenty-six. Study mode vanished overnight. No announcement. Hacker News users spotted it. Ethan Mollick at Wharton criticized it publicly, and it came back upgraded.

Which tells you what a mode even is. A behavior layer on the same brain. Originally, little more than a system prompt. Something a vendor can remove on a random Tuesday. File that away.

The current version is more substantial, to be fair. So mode one, toggle study mode on and ChatGPT stops answering and starts asking.

I tried it. Ask how photosynthesis works, and you get a question about what you already know. Then an explanation calibrated to your answer. Then a check. Can you explain it back?

And this spring's upgrades made it noticeably better. Adaptive pacing, progress tracking, and the best memory of the three. It remembers you confused mitosis and meiosis last week and quizzes you again.

Free on every plan, every region, and that continuity really compounds if you're grinding through a topic for weeks. So where's the catch?

It folds under pressure. Push back twice, "Just tell me the answer. I'm in a hurry," and it usually caves. The full answer-giving ChatGPT is one toggle away in the same window.

And eleven PM me, problem set due at midnight, pushes back twice every time. Okay, mode two. Gemini Guided Learning feels completely different.

Less like a tutor, more like a textbook chapter assembling itself around your question. Ordered chunks, diagrams generated inline, YouTube videos embedded where they help, and a quiz at the end.

The video part is almost unfair because Google owns YouTube. For first contact with something visual, organic chemistry, supply and demand curves, it's the best presentation layer of the three.

The price story turned, though. The free twelve months for students expired March eleventh, twenty twenty-six. Now it's a one-month trial, then a nine ninety-nine monthly student rate in forty-two countries.

Which now competes with a ChatGPT mode that costs nothing. And pedagogically, the gorgeous scaffold is itself the weakness?

It presents more than it asks. The quizzes arrive after the explaining, so you've been reading, not retrieving. A beautiful scaffold feels like learning, and that feeling is exactly the thing to distrust.

Mode three, Claude Learning Mode, which students describe with one word more than any other: annoying.

By design. Ask it to solve your calculus problem, and it will not solve your calculus problem. It asks what you tried, where you got stuck, what the next step might be, and it resists longest of the three.

And the institutional story is the real differentiator. Northeastern runs Claude across roughly fifty thousand students, plus the London School of Economics, Syracuse, Champlain.

It plugs into Canvas, the LMS, plus Wiley and Panopto for lecture transcripts, so the tutor lives inside the course. And Teach for All brings it to over one hundred thousand teachers across sixty-three countries.

The weaknesses follow the same philosophy. No campus deal means you get only the free toggle, none of the integrations, no progress dashboard, and the strictness genuinely grates.

The student who needs one quick factual answer late at night gets relentless questioning instead. The annoyed ones toggle it off, and then nobody is tutoring anybody.

Now the science, because this is where the three stop looking interchangeable. Start with retrieval practice. We did a whole episode on active recall, so short version.

Roediger and Karpicke, two thousand six. Students either re-studied a passage or repeatedly tested themselves. Immediately, the re-studiers looked better. A week later, it flipped hard, and the self-testers had retained substantially more.

Pulling information out of memory strengthens it in a way rereading doesn't. So for a study mode, the implication is blunt.

Every time the AI explains something you could have retrieved, it took the rep for you.

Ouch. And the second pillar got its own episode too, desirable difficulties, the Bjork research. Learning that feels harder tends to last, and a tutor that removes all struggle removes the learning along with the discomfort.

Which leaves the failure mode, cognitive offloading. Gerlach's twenty twenty-five study surveyed six hundred and sixty-six participants and found heavier AI use correlated with lower critical thinking, strongest in the youngest users.

And then the MIT team put EEG caps on essay writers. That one genuinely unsettled me.

The ChatGPT assisted group showed weaker neural connectivity and often couldn't quote from essays they had written minutes earlier. They called it cognitive debt. Borrowed performance now, paid in capability later.

Put it together and you get a five-point checklist: retrieval prompts, spacing, feedback quality, metacognition support, and escape hatch risk. And that last one may matter more than the other four combined.

Because the moment of maximum frustration is also the moment of maximum learning if you stay in it. And every product here parks a just answer me button right next to it.

So we scored all three, three points per criterion, fifteen possible. And read these loosely, they compress a lot of judgment calls. Totals?

ChatGPT nine out of fifteen. Gemini six. Claude eleven. And three findings survive any reasonable rescoring.

Let me guess the first. Claude wins on pedagogy because the thing students complain about, the refusal to just answer, is the active ingredient.

That's the entire design thesis, and the testing effect research backs it. Second, Gemini loses for the mirror image reason. Best explainer of the three. An explanation is what learning science values least.

A gorgeous scaffold you read is still rereading. And the third finding surprised me most. Nobody does spacing. Nothing in any of these products brings Tuesday's confusion back on Friday on its own.

Everything is pull, not push, and the forgetting curve does its damage between sessions. That's the open gap, and it's where we owe you a disclosure. We make Glasp, so weigh this next part accordingly.

Here's the honest framing. A study mode covers at best half of a learning system, the live session. Capture, encoding, and review on a schedule stay yours.

And the session itself evaporates into a chat log. Be honest, when did you last reread an old AI conversation?

Genuinely, never. So the fix is a capture habit you own. When something clicks, save the source. Highlight the article. Turn the lecture video into notes you keep. Pull in your Kindle underlines.

Then a few days later, interrogate instead of rereading. Ask for a quiz on your own highlights about enzyme kinetics. That's the spacing mechanism none of the modes ship.

And because it runs on your highlights, not a vendor's chat history, it survives a company quietly removing a feature on a Tuesday. We told you to file that away.

One honest limit though, the capture habit is work no tool can do for you. The encoding stays yours, and that's exactly the point.

Verdict time. No single winner, just task to tool. Walk me through the mapping.

Exam prep, where retention is everything. Claude plus your own spaced review. First contact with a hard concept, Gemini. Long self-study over weeks or months, ChatGPT because the memory and tracking compound.

Inside a university course, Claude for education if your campus has it, since nothing else has a real course system story. And if the budget is zero?

ChatGPT study mode, the only full-featured mode still free everywhere after Google's offer expired.

But the real verdict is the one I keep circling back to. The mode you pick matters less than whether you stay in it when it gets uncomfortable.

A mediocre mode used with discipline beats the best mode with the escape hatch worn smooth.

A few listener questions before we close. Is study mode actually good for exam prep?

Yes, with discipline and a supplement. Don't toggle out when it gets frustrating. That frustration is the retrieval doing its work. And add your own spaced review. The mode won't space for you.

Is using these modes cheating?

Used as designed, it's closer to hiring a tutor. It makes you do the work. Toggling it off to generate your assignment is a different activity with a different name. And your course's AI policy beats any general answer.

And which one is completely free in twenty twenty-six?

ChatGPT study mode, every plan worldwide. Claude's toggle is free, but the full campus experience needs an institutional deal. Gemini has a limited free tier, then that nine ninety-nine student rate since March.

So these 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 and compensates with the best explanations.

ChatGPT splits the difference and gives it away free. But the science doesn't move. You remember what you retrieve. You learn from difficulty you stay inside, and nobody, human or AI, can do the encoding for you. Pick the mode for the task, then build the half no mode ships.

Wanna go deeper? Full articles on glasp.co. Just search AI study modes.

See you in the next deep dive.