Custom GPTs vs Claude Skills vs Gemini Gems (2026 Guide)
Most "Custom GPTs vs Claude Skills vs Gemini Gems" posts read like feature spreadsheets written by someone who has never shipped one. I've shipped three Claude Skills in the last two months — youtube-verdict, a clickbait detector, and a video-worth-watching judge — all open source, all running in production for real users. Before that, I built half a dozen Custom GPTs and three Gemini Gems for various projects.
So this isn't a chart of checkmarks. It's the honest take after spending real time in all three.
The market context matters because all three platforms are scaling fast. ChatGPT crossed 900 million weekly active users in February 2026 (ALM Corp / OpenAI, 2026). Gemini hit 750M monthly active users by Q4 2026 (TechCrunch, 2026). Anthropic jumped to a $14B annualised run-rate in February 2026, up from $1B fourteen months earlier (SaaStr, 2026). The customisation layer on top of each of these is where most knowledge workers will actually live.
The question is which one fits your workflow. Let's do this properly.
the broader agentic AI landscape these platforms plug into
Key Takeaways
- Custom GPTs win on distribution — 900M weekly users plus a GPT Store, but you're locked to one chat product (OpenAI, 2026).
- Claude Skills win on portability — open standard, 89,753 skills published by March 2026, runs on Cursor, Codex, Gemini CLI (Simon Willison, 2026).
- Gemini Gems win on free access — went free for all users in March 2026, but ~4,000 character limit caps complexity (9to5Google, 2026).
- Decision tree at the end: pick by audience, portability needs, and how much code you actually want to write.
What Are Custom GPTs, Claude Skills, and Gemini Gems?
The three platforms solve the same problem with different philosophies. A Custom GPT is a scoped ChatGPT with custom instructions, knowledge files, and optional API actions, distributed through OpenAI's GPT Store. A Claude Skill is a markdown file (plus optional scripts) stored on your filesystem that any compatible agent can load on demand. A Gemini Gem is a saved instruction prompt with up to 10 grounding files attached, accessed inside the Gemini app.
The strategic difference is hosting. Custom GPTs and Gems live on a vendor's servers — you build them in a web UI, and they only work where that vendor lets you run them. Claude Skills live on your machine. They're literal .md files with YAML frontmatter that an agentic shell loads when triggered. That single architectural choice shapes everything downstream — portability, version control, distribution, and whether you can ever take your work elsewhere.
According to Anthropic's December 2026 announcement, Skills became an open standard with a reference SDK at agentskills.io, with launch partners including Atlassian, Canva, Cloudflare, Figma, and Notion (VentureBeat, 2026). The skill you write today on Claude Code can run tomorrow on Cursor, Codex, or Gemini CLI without changes. That's not true for Custom GPTs, and it's not true for Gems.
Why does that matter? Because the platform you build on is a bet on the platform's longevity. A Custom GPT bets on OpenAI. A Gem bets on Google. A Skill bets on the open standard — which means it works regardless of which model wins.
why the agent stack is splitting into open-vs-closed standards
How Do They Compare on Architecture and Capacity?
Custom GPTs allow roughly 8,000 characters of instructions plus up to 20 knowledge files and external API actions via OpenAPI schemas. Claude Skills have no hard instruction limit — the SKILL.md file is regular markdown, and skills can include bundled scripts that the model invokes. Gemini Gems cap at around 4,000 characters of instructions with up to 10 grounding files, no actions, no scripts (9to5Google, 2026). For complex behaviour, that capacity gap matters more than any feature checkbox.
A Gem fits a tone guide. A GPT fits a tone guide plus light tooling. A Skill fits a whole pipeline.
Custom GPTs win on integrations because actions support real OpenAPI calls — they can hit your CRM, your database, your booking system. Gems can't. Skills can, but through bundled scripts the agent runs locally rather than hosted webhooks. That difference between "hosted action" and "local script" is the entire distribution model. Hosted means anyone with a ChatGPT account can use your GPT. Local means the user has to install your skill first.
The capacity story has a simple practical implication. If your customisation is "always reply in this brand voice with these examples," any of the three works. If your customisation is "run this multi-step pipeline that fetches a transcript, scores it, verifies quotes against the source, and produces a structured report," only Skills can hold the logic without compression.
According to Bosch Research and a Carnegie Mellon study cited by Simon Willison, the Skills ecosystem grew 18.5x in twenty days — from 2,179 published skills on January 16, 2026 to 40,000+ by February 5, 2026 (Simon Willison, 2026). The capacity ceiling matters because the people pushing those numbers are doing things you can't fit in 4,000 characters.
What I Learned Building Three Production Claude Skills
I'd been building Custom GPTs for clients since GPT Store launch and had three live Gems for personal workflows. When Anthropic shipped Skills in October 2026, I rebuilt one of my Gems as a Skill to compare them properly. Then I built two more Skills from scratch — youtube-verdict, a YouTube clickbait detector, and a third that decides whether a video is worth watching. All three live in ~/.claude/skills/, all three are open source, all three solve a real problem I had.
What I actually noticed building all three: the Skill I rebuilt from a Gem doubled in usefulness because the 4,000-character ceiling had been forcing me to compress instructions until they were vague. With markdown, I could write a step-by-step pipeline, point to a verifier script, and trust the agent to follow the whole thing. The Gem version had been working at maybe 60% reliability. The Skill version cleared 95% on the same test set.
Three things became obvious once I had three Skills shipped. First, version control changed everything. Skills are files — they live in git, get reviewed, get diffed. My team can read each other's prompts. With Custom GPTs, the prompt sits in a textarea on OpenAI's site and there's no diff history. Second, scriptability solved problems prompts can't. The verifier script in youtube-verdict substring-matches every quote against the source transcript — no clever prompt makes the model verify itself reliably. With a Skill, you wire a 20-line Python script and the unreliability disappears. Third, distribution is harder than I expected. The GPT Store has 159,000 publicly listed Custom GPTs and roughly 12% of ChatGPT daily usage flows through one (SEO.ai, 2026). Skills don't have an equivalent storefront yet — Vercel's skills.sh marketplace listed 89,753 skills as of March 2026 (Simon Willison, 2026), but discovery is still through GitHub README files and word of mouth.
The painful part of Skills? No GUI builder. Nothing to click. You write markdown and YAML frontmatter and test in your terminal. For me that's a feature — I'd rather edit a file than poke a web UI — but if you've never written markdown, the on-ramp is steeper than ChatGPT's "Create GPT" button.
Building three Skills in two months convinced me the architecture is right but the polish lags ChatGPT by about a year. The capability ceiling is much higher; the developer experience for non-coders is much lower. That gap closes if Anthropic ships a builder UI in 2026 — and given the trajectory, they probably will.
the architecture and design decisions behind the youtube-verdict skill in detail
How Do Custom GPTs Compare in 2026?
Custom GPTs are still the easiest to build and the easiest to distribute. ChatGPT crossed 900 million weekly active users in February 2026, with Custom GPTs and Projects making up roughly 12% of daily usage and weekly users of those features growing 19x year-to-date (Digital Applied / OpenAI State of Enterprise AI, 2026). 92% of the Fortune 500 use ChatGPT in some capacity (Christian & Timbers, 2026). That distribution is unmatched, and it's the single best reason to build a Custom GPT instead of a Skill.
Mass reach versus enterprise spend — the two platforms are optimising for different ends of the curve.
The GPT Store reality is messier than it looks. There are 3 million+ Custom GPTs created and roughly 159,000 publicly listed (SEO.ai, 2026), which means the discovery problem is brutal — your GPT competes with thousands of similar ones. Revenue sharing exists but pays out at roughly $0.03 per conversation, so unless your GPT goes viral, it's a portfolio piece rather than a business. ChatGPT pricing as of April 2026 is Free, Go ($8), Plus ($20), Pro ($100, launched April 9, 2026), Pro+ ($200), and Business at $20-30 per seat (IntuitionLabs, 2026). Custom GPT actions require Plus or above, so the free tier doesn't get tools.
Custom GPTs make sense in three specific cases. You're targeting non-technical users who already use ChatGPT — that's most of your office. You need a hosted action that hits a public API. You want a portfolio piece in the GPT Store for inbound leads. Outside those cases, you'll hit the same wall I did: it's easier to ship, harder to evolve, and impossible to take with you.
why ChatGPT Projects and Custom GPTs together still won't replace specialist tooling
Where Do Gemini Gems Fit?
Gems went free for all Gemini users on Android, iOS, and web in March 2026, after starting as a Gemini Advanced perk (Tom's Guide, 2026). Free is the headline feature. The Gemini app crossed 750 million monthly active users by Q4 2026, AI Mode hit 75 million daily users, and Gemini Enterprise has 8 million paid seats across 2,800+ companies (TechCrunch, 2026). If your audience uses Android, they almost certainly have Gemini installed.
The Gem itself is the lightest of the three. A name, a description, an instruction prompt of roughly 4,000 characters, and up to 10 attached files for grounding. No actions. No scripts. No external API calls. The mental model is "saved chatbot persona," and that's both its strength and its ceiling. For a personal writing assistant, a daily standup helper, or a kids' homework tutor — all of which I've built — Gems are excellent. The friction is so low that I built one in three minutes the first time I tried.
For anything beyond a persona, Gems run out of room fast. The 4,000-character cap means you can describe behaviour but you can't enumerate edge cases. You can attach a brand guide as a grounding file, but you can't run a verification step. The first Gem I built ("clickbait detector") rewrote itself as a Skill within a week because I needed multi-step reasoning that wouldn't fit. That's the recurring pattern with Gems — they're a 90th-percentile-easy entry point that runs into a low ceiling for serious work.
If you're a content creator with a phone and an idea, Gems are unbeatable. If you're a developer trying to ship a tool, Gems are training wheels you'll outgrow in a month. The free tier is the right move from Google — it's also a reminder of where Google sees Gems on the value chain.
The other thing worth noting is that Gems are tightly bound to the Gemini app. There's no Gem CLI, no Gem-on-Cursor, no portable Gem format. Whatever you build only runs where Google lets it run. For some buyers that's fine — for builders chasing portability, it's the deal-breaker.
When Should You Pick Each Platform? The Decision Tree
The honest answer is "it depends on your audience, your portability needs, and how much code you want to write." Most comparison posts dodge that complexity with a feature table. I'd rather give you an actual decision tree.
A four-question decision tree. Portability and audience are the two questions that matter most.
Walk it through with the four personas I've built for. The solo developer building tooling for themselves and their team picks Claude Skills — version control, scriptability, runs on Cursor and Codex, no vendor lock-in. The agency selling AI helpers to non-technical clients picks Custom GPTs — those clients already have ChatGPT, the GPT Store gives a discoverability story, and OpenAPI actions cover most integration needs. The content creator on a phone wanting a brand-voice helper picks Gemini Gems — it's free, takes three minutes to build, and runs anywhere Gemini does. The enterprise rolling out a custom AI assistant to 5,000 employees? That depends on which model their security team has already approved, but increasingly that's Claude — 1,000+ enterprises now spend over $1M per year on Claude, doubled from 500 just two months earlier (DemandSage, 2026).
From my own decision matrix: of the seven AI helpers I've built in 2026, five became Skills, one stayed a Gem, and one is a Custom GPT. The Skill-versus-Gem split tracked exactly with whether the helper needed a verifier script. The Custom GPT survived because the audience for that one specific tool — non-technical marketers at a client agency — was 100% on ChatGPT and zero percent on Claude.
The decision tree breaks down in two edge cases. If your model preference is fixed for safety or compliance reasons, the tree doesn't matter — you build wherever your IT team lets you. If you're optimising for monetisation rather than capability, the GPT Store revenue share, however small, is the only one of the three that pays anything at all today.
the broader question of when chatbot wrappers actually replace human work
What's the Verdict for 2026?
If I had to compress my view into one sentence: build on Claude Skills if you're a developer, build on Custom GPTs if you're shipping to non-technical users on ChatGPT, and build on Gemini Gems if you need free distribution to a phone audience. The platforms aren't really competing for the same buyer — they're each optimised for a different point on the build-versus-distribute curve.
Where I think this is heading: the open-standard bet is the most under-priced part of this comparison. Skills runs on Cursor, Codex, Gemini CLI, and 30+ other tools as of early 2026. That means Anthropic's customisation surface area extends well beyond Claude itself — every agent-shell that adopts the standard becomes a free distribution channel. Compare that to Custom GPTs, which only run on ChatGPT, and Gems, which only run on Gemini. The platform that decoupled customisation from the model is the platform whose work survives a model shake-up.
The numbers back the trajectory. Anthropic doubled its enterprise high-spend cohort in two months and crossed a $30B annualised run-rate by April 2026 (SaaStr, 2026). Skills crossed 89,753 published artifacts on skills.sh by March 2026 (Simon Willison, 2026). Custom GPTs are still a much larger ecosystem in raw count, but the growth curve has flattened. Gems remain a consumer feature with no obvious path to a developer ecosystem.
What about a year from now? My guess: Custom GPTs become a hosted distribution channel for marketing-style helpers, Skills become the dominant developer-facing format, and Gems keep doing what they do well — letting normal humans save a chatbot persona without learning anything new. All three survive because they target different buyers. The interesting question isn't which one wins. It's which one your work is portable to when the underlying model changes.
Frequently Asked Questions
Are Custom GPTs still worth building in 2026?
Yes — if your audience is on ChatGPT. With 900M weekly active users and Custom GPTs accounting for 12% of daily ChatGPT usage (SEO.ai, 2026), the distribution is unmatched. They're less compelling for builders who need portability across agents, since GPTs only run inside ChatGPT.
Can Claude Skills work outside Claude?
Yes. Skills became an open standard in December 2026 with a reference SDK at agentskills.io (VentureBeat, 2026). They run on Cursor, Codex, Gemini CLI, and 30+ other agentic shells that have adopted the spec. One skill works across every tool you use — the open-standard bet is Anthropic's strongest moat.
Do Gemini Gems support tool calls or actions?
No. Gems are limited to instruction prompts (around 4,000 characters) and up to 10 grounding files for context. They cannot make API calls, run scripts, or execute multi-step pipelines. For anything beyond a persona or tone helper, you'll hit the ceiling fast and need to graduate to Claude Skills or Custom GPTs.
Which platform has the best free tier?
Gemini Gems went free for all users in March 2026 (Tom's Guide, 2026) — no subscription needed on Android, iOS, or web. Claude Skills are free to author but need a Claude account to run ($20/month Pro tier). Custom GPTs are free to use but actions require ChatGPT Plus ($20+/month).
Will any of these still matter in two years?
Probably yes for all three, but for different reasons. Custom GPTs ride ChatGPT's distribution. Gems ride Gemini's mobile ubiquity. Claude Skills ride the open-standard bet — they outlast any single model. The platform whose work travels with you when the model landscape shifts is the one with the longest lifespan, and that's currently Skills.
Conclusion
I built three Skills in two months and shipped Custom GPTs and Gems before that. The honest takeaway from doing the work, not just reading the docs: the three platforms aren't really comparable on a feature checklist because they're solving different parts of the same problem.
Custom GPTs are the easiest path to non-technical users. Gemini Gems are the easiest path to a free, mobile audience. Claude Skills are the only path that survives if the model underneath changes. Pick by audience and portability, not by feature spreadsheet.
If you're a builder, my bias is obvious — Skills won me over because they're files, files go in git, git makes work durable. If you're not a builder, build a Custom GPT or a Gem and don't apologise for it. The right platform is the one your audience already uses.
Top comments (0)