What teachers were really building with custom GPTs — and what comes next

Custom GPTs are going away. OpenAI announced in September that they're retiring the feature across all plans, with a hard deadline of December 11, 2026.

If you're an educator who built custom GPTs for your classroom, you already know this. You probably got the in-app notice. You may have already clicked "Migrate to plugin" and discovered that your sharing settings didn't carry over, your students lost access, and the thing you spent hours building now sits in a private folder that only you can see.

I've been following the community reaction closely, and the frustration is real — particularly from teachers. A University of Florida instructor asked OpenAI to preserve a no-code path for sharing course chatbots with students. A researcher described feeling "rather foolish for having recommended a platform feature immediately before OpenAI withdrew the very capability that made it useful." Independent creators who built educational GPT tools are asking how they're supposed to serve customers who already paid.

The advice floating around is to switch to ChatGPT Projects, Claude Projects, or Google Gemini Gems. These are fine tools. But I think the conversation is missing something.

What were teachers actually building?

When a teacher created a custom GPT, they weren't just making a chatbot. They were trying to build something specific: an AI assistant that understood their course materials, responded in a way appropriate for their students, and could be shared via a simple link.

That's a real need. And custom GPTs were the closest thing to meeting it — until they weren't.

The problem is that custom GPTs were always a workaround. Teachers were taking a general-purpose consumer tool and layering on instructions, uploaded documents, and conversation starters to approximate something that should have been purpose-built. There were no pedagogical guardrails. No way to prevent the AI from just doing the student's work. No standards alignment. No grading. No moderation system appropriate for minors. No visibility into what students were actually doing or learning.

Teachers accepted these tradeoffs because custom GPTs were free, easy to set up, and shareable. Now that they're gone, the question isn't just "what replaces them?" It's "what should we have been building toward all along?"

The alternatives list is missing the point

Here's what most comparison articles recommend as replacements:

• ChatGPT Projects: Custom instructions + reference files + saved conversations. Private to you.
• Claude Projects: Similar concept. Also private.
• Gemini Gems: Closest to the original custom GPT experience within Google's ecosystem.
• Copilot Studio: For schools already on Microsoft 365.
• Third-party builders (Pickaxe, Chatbase, Dify): Roll your own chatbot.

These all solve the "chatbot with custom instructions" problem. None of them solve the "students actually learning" problem.

A teacher uploading a PDF to a Claude Project still gets a general-purpose AI that will happily write a student's essay for them. There are no guardrails that ensure the AI asks questions instead of giving answers. There's no way to see whether a student engaged meaningfully or just copy-pasted a prompt. There's no alignment to learning standards. There's no grading rubric. There's no moderation system flagging concerning content in real time.

These aren't edge cases. They're the core requirements of using AI with students.

What it looks like when you design for education from the start

At Flint, this is what we've spent the last three years building — not a chatbot with instructions bolted on, but a personalized learning platform designed around how teachers teach and how students learn.

When a teacher creates an activity on Flint, they get everything a custom GPT offered (custom instructions, uploaded materials, a link for students) plus everything it didn't:

The AI teaches instead of answering. Flint's AI, Sparky, uses Socratic questioning and scaffolding to guide students through problems. It creates productive struggle rather than bypassing it. This is the single biggest gap in every general-purpose AI tool — they're built to be helpful, which in a classroom context means doing the student's thinking for them.

Standards alignment is built in. Teachers link activities to state standards and learning objectives. Sparky tracks mastery per student, per standard, across sessions. A teacher can see which standards a class has mastered and where gaps remain. No general-purpose AI tool knows what the Common Core or Next Generation Science Standards are, let alone tracks student progress against them.

Grading works. Teachers define rubrics. Sparky grades based on those rubrics, adapted to individual student progress. Teachers can review, override, and add their own feedback. This isn't an afterthought — it's integrated into the activity flow.

Safety and moderation are real. Content is flagged in real time against configurable categories. Self-harm alerts go out immediately. Schools set their own moderation policies, assign moderators, and choose alert cadences (immediate, hourly, daily). When your students are mostly minors, this isn't optional.

Sharing actually works. Teachers create an activity. Students access it in their class. There are no broken links, no "start private and manually re-share," no wondering whether a student can see the thing you built. It's part of the class structure, integrated with SIS/LMS roster sync.

Stability matters

One thing that kept coming up in the educator reactions to the custom GPT deprecation: teachers need platforms that won't yank features mid-year. Tina Austin, writing on Substack, put it plainly — educators are looking for stable platforms because planning a semester around a tool that might disappear is untenable.

This is something we take seriously. When a school partners with Flint, they're adopting a platform built specifically for their context — not a consumer feature that might get deprecated when the company's product strategy shifts. We co-build with schools. Our roadmap is shaped by what educators tell us they need, not by what makes a good demo at a developer conference.

Where this is heading

Custom GPTs gave teachers a taste of what's possible: AI that's customized for their classroom. The fact that so many educators invested time building them — despite the lack of pedagogical guardrails, safety features, or student visibility — shows how strong the demand is.

The right response to the deprecation isn't to rebuild the same chatbot on a different platform. It's to ask what teachers were actually trying to accomplish, and to give them a real tool for it.

If you've been using custom GPTs with your students and you're figuring out what's next, I'd encourage you to try Flint. We have a free tier for individual teachers, and we'd genuinely love your feedback on how we can make it better.

— Sohan