David E. Jacobson | History Teacher
Last month, Claude Sonnet 5 dropped. Two weeks before that, someone in a teacher Slack I’m in asked whether they should switch from Gemini to ChatGPT-4 because they read that it was better for grading. A colleague sent me a three-minute demo of a new AI grading tool that looked slick. A parent asked if I was using the latest version of MagicSchool.
I didn’t make a single change.
That’s not because I’m not curious. I read about the releases. I think about what new features might mean for how I give feedback or build activities. But I’ve also learned something over the last couple years of living in the AI conversation: the treadmill of new models is real, and it’s not helping your students. What helps your students is you knowing your tools so deeply that you can get out of the way and let the pedagogy do the work.
The Churn Is Relentless
Here’s what’s happening in the AI landscape right now. Between February and July 2026, Anthropic released four major Claude updates: Opus 4.6, 4.7, 4.8, and now Sonnet 5. OpenAI has GPT-4 variants. Google has Gemini. Each release comes with benchmarks claiming this one’s better for reasoning, this one’s faster, this one costs less. For education specifically, there’s a growing suite of tools built on top of these models — MagicSchool, Flintk12, Granola, custom bots, each with their own updates and integrations.
If you’re a teacher trying to stay current, it’s an actual treadmill. Each one promising to make your life better. Each one requiring you to relearn a workflow, adjust your prompts, sometimes rebuild an entire activity around the new capabilities.
When I looked at the data, though, something stood out: it’s not working. A 2025-2026 study across educator communities found that as the number of AI tools available to teachers tripled, the percentage of students who felt supported by AI-assisted learning dropped from 29% to 13%. The practitioner consensus was brutal but clear: stop adding tools. Get very good at a very short list.
What Consistency Does
Your students have enough variables in their lives already. They’re navigating a new class, new classmates, new expectations for how they think and write. Their anxiety about AI—whether they should use it, whether their teacher will think they cheated, whether they’re lazy if they lean on it—is already in the room.
When you introduce a new tool or swap out the model you were using, you’re adding another variable. They have to relearn the interface. The prompt structure that worked last week doesn’t work the same way. They’re confused about what the AI is doing differently, and you’re the one who has to re-explain it.
Worse: you haven’t built the mental model yet for how this new tool behaves. You’re learning alongside your students. That sounds collaborative, but it misses the point—you’re unprepared.
I spent probably two months last year getting good with Claude for giving feedback on essays. I understood the rubric I needed to feed it, the way to structure the prompt so it wouldn’t miss certain subtleties about a student’s argument, how to handle student names and context so the feedback stayed personal. By the end of those two months, I wasn’t thinking about Claude anymore. I was thinking about the essay. The tool had become invisible.
Six months in, I could give better feedback faster and with more consistency across 120 papers than I could when I was still figuring out the tool. What students experience is consistency, not novelty—it’s the reliability that comes from a teacher who knows her tools.
The Real Place to Experiment
Here’s where the new models matter: outside the classroom, on your own time, when you’re not managing 30 students and a lesson plan.
When Claude Sonnet 5 dropped, I spent an evening testing it. I ran some of my standard prompts against it and compared the output to what Opus 4.8 was giving me. I played with a new feature I’d read about. I broke things on purpose to see how it would respond.
Did I find anything that would change what I do in the classroom? Not really. Sonnet 5 is faster and cheaper, which is nice. But the essays I’m grading, the activities I’m building, the way I’m thinking about AI in the classroom—none of that changes because there’s a marginally better model available.
If I had found something that would make a meaningful difference—say, a feature that finally let me do exactly what I wanted with differentiated scaffolding—then I would experiment a little more. I’d rebuild one activity. I’d test it. I’d see if it landed differently with students than what I was already doing. And then, if it was worth it, I’d commit to changing the whole workflow.
But I wouldn’t do that during a unit where my students are in the middle of learning. I wouldn’t flip the model mid-semester just because something new came out. That’s not experimenting. That’s disrupting.
The Cost of Switching
There’s a real cost to switching models or tools in a classroom environment, and it’s not just the time you spend relearning.
It’s the loss of accumulated knowledge. All those months of understanding how a tool behaves, what it’s good at, what it misses—that knowledge is gone when you switch. You’re back to being the one asking “wait, why did it do that?” while your students are counting on you to know.
It’s the inconsistency students experience. Now the feedback they get from you is generated by a different model. It has a slightly different voice, different priorities, different blind spots. They notice. They start second-guessing whether what they learned from last month’s feedback still applies.
It’s the precedent it sets. If you’re swapping tools every time something new and shiny comes out, you’re modeling that instability. You’re saying “the tool matters more than the work.” Your students learn that too.
The One Exception
If your current tool is failing you—because it’s not available anymore, because it’s worse at the thing you need it for, because your school blocks it and you have to find an alternative—then yes, switch. Find something that works, get good at it, and stay with it.
But if your tool is doing the job, if your students are learning, if you’ve built competence with it, then the bar for switching shouldn’t be ‘there’s a new release.’ It should be ‘this new tool will materially change what’s possible in my classroom, and I’ve verified that on my own time.’
My Rhythm
Here’s what I actually do: I have a tool that’s working for me in the classroom. I know it well. I use it intentionally. That’s the constant.
Outside of that, on my own time, I stay curious. I read about new releases. I think about what new features might mean for how I give feedback or build activities. But I keep that separate from what my students experience. I don’t drag my classroom into my experimentation.
And when I find something worth switching to—which I might, eventually—I’ll do it the right way. I’ll build my competence first. I’ll test it outside the classroom. I’ll make sure it’s not just novelty but an improvement. And then I’ll commit to it without second-guessing every time the next new model drops.
The truth is, most of these releases don’t matter to what happens in your classroom. A model that’s 2% faster at grading doesn’t change your pedagogy. A new feature you’ll use twice a year isn’t worth retraining yourself on. And your students need consistency more than they need you to be on the bleeding edge.
Stay curious. Experiment on your own time. But keep the classroom steady. That’s where the real work is.
What’s your approach? Are you experimenting with new models, or have you found something that works and you’re sticking with it? I’d love to hear how other teachers are navigating this.
If this tension between consistency and experimentation is something you’re thinking through, my book goes deeper. The AI Doesn’t Know Your Students is available on Amazon and at shouldiuse.ai/book.
