Gems vs Custom GPTs vs Projects: Reusable AI Setups Compared

TL;DR: Gems, Custom GPTs, and Claude Projects all let you save instructions and files so you stop retyping context every session, but none of them travels with you. Custom GPTs can no longer be created on any personal ChatGPT account. Claude's own help center contradicts itself on who gets RAG. And every one of the three is locked to its own vendor's model, which is the actual decision this comparison comes down to.

What Are Gems, Custom GPTs, and Projects, in One Line Each?

Gems (Google) are, per Gemini's own overview page, "your custom AI experts for help on any topic". They're built from saved instructions plus optional uploaded files, usable inside Gemini Apps, the Gemini side panel in Google Workspace, and the Gemini mobile app.

Custom GPTs (OpenAI) are ChatGPT's equivalent: a saved system prompt plus optional files and Actions, historically the best-known version of this pattern, going back to November 2023.

Projects (Anthropic) group Claude conversations under shared instructions and uploaded "project knowledge", per Claude's own help center, with a newer auto-summarizing memory layer that carries context across a project's chats.

All three solve the same basic problem: stop retyping the same role, tone, and background context every time you open a new chat. Where they differ, sharply, is who can actually build one, and what happens the moment you want to use that setup somewhere other than the vendor that hosts it.

Who Can Actually Create One?

This is the single biggest, least-known difference among the three, and it's worth leading with rather than burying.

Custom GPTs are the most restricted of the three, and it's a recent, easy-to-miss change. OpenAI's own help center states it plainly, in two separate articles read directly on September 5, 2026: "New GPT creation and publishing are not available on personal ChatGPT accounts, including Free, Go, Plus, and Pro." Creation now lives inside "Business, Enterprise, and Edu workspaces", and even there, per the same page, GPT creation, editing, and publishing "depend on workspace settings and permissions." If you're a solo user on a personal Plus subscription, you cannot build a new Custom GPT at all, full stop, regardless of how much you're paying OpenAI.

Gems have the lowest bar of the three. Google's own support pages state only an age requirement: "you must also be 13 (or the applicable age in your country) or over." We could not find a stated plan requirement, free-tier Gem count, or workspace restriction anywhere on Google's own pages we checked.

Projects sit in the middle, with a documented free-tier cap and an internal contradiction. Anthropic's own help center states: "Projects are available to all users, including those with free Claude accounts. Free users can create a maximum of five projects." That same article, further down, calls using project instructions something you do only "when you need specific guidance or context for a particular project (paid plans only)." Both sentences are live on the same page; we're reporting the contradiction rather than picking a side.

Worth sitting with for a second: OpenAI's restriction is the one most people evaluating this comparison won't already know about, because it's a genuinely recent change and a huge amount of still-ranking content on the open web assumes any Plus subscriber can build a Custom GPT the way that was true when the feature launched in November 2023. If you've been putting off building one because you assumed you needed a higher personal tier, the actual bar has moved past any personal tier entirely; it's now a workspace-account requirement, which is a different kind of purchase than a subscription upgrade.

Before you invest hours in any of the three:
[ ] Confirm your account type can even create one (Custom GPTs: workspace only)
[ ] Check the CURRENT capacity limit on the vendor's own page, not a cached figure
[ ] Decide if "reusable" means "reusable in this one app" or "reusable everywhere"
[ ] If multi-model matters, price a vendor-neutral library against the rebuild cost

How Much Can Each One Actually Hold?

Capacity is the second axis worth checking before you invest hours building any of these, and here the picture is genuinely uneven across vendors.

Gems let you attach files under a section Google's own support page labels "Knowledge", with three separate upload paths: from your own device, from Google Drive, or from a NotebookLM notebook. One detail worth knowing if you edit source files after attaching them: "If you add a file from your Drive, Gemini will use the most recent version of the file." That means a Gem built against a Drive doc stays current automatically rather than freezing at attachment time.

Custom GPTs publish no character limit for their Instructions field at all, on any OpenAI page we could find. A widely-repeated 8,000-character figure some sites still cite doesn't appear anywhere in OpenAI's current documentation; treat it as folklore, not a stated spec.

Projects publish the most detail of the three, and also the most self-contradictory detail. Anthropic's RAG documentation states project knowledge can, once RAG activates, expand "capacity by up to 10x while maintaining quality responses." But whether RAG is even available to you depends on which of two contradicting Anthropic pages you read: one states RAG "is only available to users with paid Claude plans (Pro, Max, Team, or Enterprise)"; a separate, also-current page states RAG "is available for all Claude plans (free, Pro, Max, Team, and Enterprise)." Both were live when we checked. We are not resolving that for you; check both pages yourself before assuming either applies to your account.

Gems, Custom GPTs, and Projects, Side by Side

What Happens When You Switch Models?

This is the question that actually decides whether any of these three is worth building out, and none of the three vendors answers it in your favor.

A Gem runs on Gemini's models and nowhere else. A Custom GPT runs on OpenAI's models and nowhere else, assuming you're even in a workspace that can build one. A Claude Project runs on Claude and nowhere else. None of the three publishes an export format that turns your saved instructions and files into something another vendor's equivalent feature can import directly.

That means every hour you spend perfecting a Gem's instructions is an hour of work that stays inside Google's product. Standardize your team's onboarding assistant as a Custom GPT, and the moment leadership asks whether you should also try Claude for this, you're rebuilding it as a Project from scratch, assuming your account can even create one. This isn't a criticism unique to any one vendor; it's the structural consequence of building a reusable setup inside a single chat provider's own walled feature, whichever provider you pick.

Walk through what actually gets lost in that rebuild, because it's more than just retyping the same paragraph. A Gem's uploaded Drive files, a Custom GPT's Actions configuration and knowledge files, a Project's uploaded documents and any RAG-expanded knowledge base: none of that transfers as a file you can export and re-import elsewhere. You're not just rewriting instructions; you're re-uploading source material, re-testing edge cases the original setup had already caught, and re-learning whatever quirks the new vendor's version of the feature has that the old one didn't. For a single personal Gem, that's an afternoon of annoyance. For a team's shared onboarding assistant, client-facing support setup, or documentation Project that took months of iteration, that's a real cost worth pricing out before you commit hard to any one vendor's walled version of this feature.

It's also worth being honest about why none of the three vendors builds an export path. A Custom GPT that could be exported and run identically on Claude or Gemini would weaken exactly the reason OpenAI wants you building inside ChatGPT in the first place, and the same incentive applies in reverse for Google and Anthropic. Lock-in here isn't an oversight any of the three is likely to fix voluntarily; it's a fairly direct consequence of each vendor wanting your best work to stay inside its own product.

Is There a Reusable Setup That Isn't Locked to One Vendor?

The honest answer is: something has to sit outside all three products if you actually want portability, and that's a real market gap all three vendors leave open rather than one any of them is trying to fill.

A saved prompt library, kept independently of any single chat provider, is the more portable version of the same idea. Ours works this way deliberately: Prompt Architects' Personal Context feature is attached manually to a prompt, one toggle switch away, rather than auto-injected into every conversation, and each saved context holds a flat 5,000-character budget. That's a smaller per-context allowance than what Claude's RAG-expanded Projects can reach, but it works identically whether you're generating a prompt for ChatGPT, Claude, or Gemini, which none of the three tools above can claim about each other. We cover the same underlying tension, reusable AI setups versus vendor lock-in, from the provider-settings angle in Provider Settings Compared: OpenAI vs Anthropic vs Google, and from Google's own alternatives in Gemini Gems Alternatives for Reusable AI Setups.

None of this makes Gems, Custom GPTs, or Projects a bad choice for staying entirely inside one vendor's ecosystem; if your team only ever uses Claude, Anthropic's Projects are a reasonable, well-documented place to build shared context, contradictions in its own docs aside. The choice to make deliberately is whether reusable means reusable inside one vendor, or reusable everywhere you actually work, because the three tools compared here only promise the first one. If your team is weighing a prompt library against Custom GPTs specifically, from the library side of that argument, Custom GPTs vs a Prompt Library: Which Scales Better? goes deeper on exactly that trade-off.

So, Which Should You Use?

If you only ever use one model and have no plan to change, pick that vendor's own feature and don't overthink it: a Gem if you're on Gemini, a Custom GPT if your workspace can create one, a Project if you're on Claude. Each is well-integrated with its own ecosystem, and none of the cross-vendor portability concerns above matter if you're never going to test the alternative.

If there's any real chance your team evaluates a different model in the next year, or already uses more than one, the calculus changes. Building your best reusable instructions and context inside a single vendor's walled feature means redoing that work the day you switch, and given how fast model landscapes have moved through 2026, betting your whole setup on never switching is harder to justify than it was even a year ago. A model-agnostic prompt library, ours or a competitor's, at least survives that switch intact.

There's also a middle path worth naming rather than skipping past: plenty of teams genuinely do use more than one model already, on purpose, for different jobs, without wanting to standardize on a single vendor at all. If that's your actual situation, none of Gems, Custom GPTs, or Projects was ever going to be a single answer for you, since each covers exactly one of the three models you're already switching between. That's the clearest case for a vendor-neutral layer sitting above all three: not a bet against any one of them being good at the one model it runs on, but an acknowledgment that being good at one model and being the setup your whole workflow depends on are different requirements once more than one model is genuinely in play.

The failure mode worth avoiding either way is discovering the lock-in after you've already sunk the hours in. A team that spends a quarter perfecting a Custom GPT, only to have a workspace policy change or a new hire push for Claude, learns this lesson the expensive way. Naming the trade-off before you start building, rather than after, costs nothing and saves exactly that outcome.

The most useful thing this comparison can tell you isn't which of the three is "best." It's that all three are the same shape of bet: convenient inside one vendor's product, and a rebuild the moment you leave it. Decide with that trade-off named plainly, rather than discovering it the day you actually need to switch.