Is Notion AI using ChatGPT?

Why Nobody Can Tell What Powers Notion AI

Notion markets AI everywhere — headlines, landing pages, feature demos — yet the one place you'd expect a straight answer, its own product page, never names a single model provider. Visit Notion's official AI page and you'll find feature descriptions without a whisper of OpenAI, ChatGPT, GPT, or Anthropic anywhere in the copy.

That silence forces users into detective work. The truth about Notion AI's model stack lives almost entirely in third-party blogs and partner case studies, not in official Notion documentation. Independent sources fill the gap: Notion AI launched in November 2022 as "just a GPT model integrated in Notion blocks," then added Anthropic's Claude in March 2024, according to a detailed timeline from a Notion consultant. None of that comes from Notion itself.

The picture those sources paint is deliberately fuzzy by design:

• Notion AI reportedly routes tasks automatically across multiple models — GPT for creative writing, Claude for long documents — though the source describing this routing names no specific versions and cites no primary documentation.
• The strongest recent OpenAI connection — a November 2025 case study describing Notion rebuilding its agent system on GPT-5 — survives only secondhand via a Notion partner's blog.
• No source specifies which exact model version powers which feature, and Notion has never published a technical audit of its architecture.

Why does this opacity matter? Because since May 13, 2025, you can no longer test Notion AI as a cheap add-on. It ships only with Business and Enterprise plans at €19.50 per seat per month — a real budget line for teams committing before knowing what they're actually paying for. And because Notion AI uses the same OpenAI LLMs behind ChatGPT, it inherits the same weaknesses: hallucination, knowledge cutoffs, and token limits. If you're evaluating it against ChatGPT or Claude directly, the model underneath is part of the value calculation.

This is a familiar pattern to us at SkillzBench: the AI ecosystem routinely asks users to trust tools whose internals nobody discloses. Our own answer to that problem — version-pinned, sandboxed benchmarks that measure what a component actually does rather than what marketing claims — is still in development, but the principle applies here. When a vendor won't tell you what's under the hood, independent verification is the only substitute.

So the core question — is Notion AI using ChatGPT? — has a real answer, but you won't find it on Notion's website. You'll find it by piecing together third-party evidence, and the answer is more nuanced than a simple yes or no. Here's what that evidence actually shows.

The Answer: Yes, But Not Exclusively ChatGPT

The short answer is yes — but the full story is more interesting than a simple yes. Notion AI runs on OpenAI's models, yet calling it "ChatGPT in a workspace" misses how the product actually works today.

Notion AI launched in November 2022, "right when ChatGPT became popular," and according to independent Notion consultant documentation, it "started out to be just a GPT model integrated in Notion blocks." That GPT foundation remains, and the OpenAI relationship appears to be deepening: an OpenAI case study cited in November 2025 describes Notion rebuilding its agent system on GPT-5, moving "beyond isolated prompts toward an architecture designed for reasoning models and tool orchestration."

But OpenAI is no longer the only player. In March 2024, Notion added Anthropic's Claude to the mix, and reporting indicates tasks are routed automatically to the best-suited model — GPT for creative writing and brainstorming, Claude for long documents and detailed analysis. Some features even let users pick a preferred model, while many requests are handled behind the scenes.

The key distinction is this: Notion AI uses OpenAI's models, but it is not ChatGPT. As one analysis puts it, "Notion can use OpenAI models inside its AI features, but they're not the same product." Notion is the workspace and product experience; OpenAI provides the underlying intelligence. Notion develops no LLM of its own — it integrates third-party models so they can work inside your pages, databases, and connected tools like Slack and Google Drive.

Think of it like apps running on an operating system:

• The LLMs (GPT, Claude) are the OS-level intelligence layer
• Notion AI is an application layer that adds workspace context, permissions, and pre-built prompts
• ChatGPT is a different app on the same OS — a conversational interface without your workspace data

Notably, Notion's own product page never mentions OpenAI, ChatGPT, GPT, or Anthropic — the model stack is only confirmed by third-party sources. That opacity is exactly the kind of transparency gap SkillzBench is building verification and benchmarking infrastructure to close for AI agent skills. The practical takeaway: Notion AI shares ChatGPT's underlying strengths and weaknesses — including hallucination and knowledge cutoffs — while its real differentiator is that it's grounded in your workspace, not the open web.

What This Means in Practice: Capabilities and Shared Weaknesses

Knowing that Notion AI runs on GPT and Claude models only matters if it changes how you use the tool — and it does, in both directions.

In practice, the multi-model architecture means Notion acts like a dispatcher. According to reports on Notion's model stack, GPT-style models handle creative work — blog drafts, marketing copy, email drafts, and outlines — while Claude takes on long-document analysis like report summaries, meeting transcripts, and knowledge bases. Some features let you pick a model; many requests route automatically.

The real differentiator isn't the models at all. Notion AI's Q&A is grounded in your workspace — it answers from content you can access, respects workspace permissions, and pulls in connected tools like Slack and Google Drive rather than just the current page, per independent analysis. As one review puts it, the Q&A feature "does not access the web, retrieve current information, or analyze sources that live outside of Notion."

That grounding cuts both ways. You get answers tied to your actual pages and permissions, but you don't get current web information. Notion AI improves how you navigate and act on content you've already written — it doesn't transform how you capture it.

The flip side is inherited weakness. Because Notion AI uses OpenAI's LLMs — the same models behind ChatGPT — it shares their known limitations:

• Hallucination: the model can state things confidently that aren't accurate
• Knowledge cutoff: no access to information newer than the model's training data
• Token limits: very long documents may exceed what the model can process at once

One caveat deserves honesty: Notion's own product page discloses none of this. It describes AI features without mentioning OpenAI, Anthropic, or any model provider. The model claims in this article come from third-party sources, and even the strongest recent evidence — a GPT-5 agent rebuild reported in November 2025 — is secondhand, citing an OpenAI case study rather than a Notion statement. Exact model versions per feature remain unspecified.

This opacity isn't unique to Notion — it's the norm for AI products, and it's exactly the trust gap SkillzBench is building infrastructure to close for AI agent skills: version pinning, security scanning, and sandboxed Lift benchmarks, so capability claims get verified rather than assumed. Until that kind of verification is standard, treat any AI tool's model claims — including these — as well-sourced inference, not vendor-confirmed fact.

The Transparency Gap — and How SkillzBench Is Closing It

Here's the thing: if you want to know which AI models power Notion AI, you won't find the answer on Notion's own product page. The company describes its AI features in detail but never mentions OpenAI, Anthropic, GPT, or Claude. Everything we know about Notion AI's model stack comes from third-party sources piecing the picture together.

That should give every AI buyer pause. Notion is a mainstream product used by millions, yet its AI architecture is effectively opaque. The GPT-5 claim — that Notion rebuilt its agent system on OpenAI's latest model in November 2025 — is secondhand, cited via an OpenAI case study rather than any Notion documentation. Even the multi-model routing story, where GPT handles creative writing and Claude handles long documents, comes from blogs without primary citations.

This is the transparency gap in miniature, and it's not unique to Notion. It's the default state of the AI ecosystem.

Agent skills — the SKILL.md instruction packages that power Claude Code, Codex, Cursor, and Gemini CLI — face an even sharper version of this problem. Published research found roughly 26% of skills carry at least one vulnerability. Nobody can tell at a glance which skills actually help, which are risky, or what changed since last week.

SkillzBench, an independent testing lab from AIQ Labs, is building the infrastructure this ecosystem lacks:

• Version pinning — every skill version is pinned to a specific commit SHA, so you always know exactly what you're running
• Security scanning — automated vulnerability, license, and quality checks on every listed skill
• Sandboxed Lift benchmarks — each skill is tested with the same model, with and without the skill, and the performance Lift is published in percentage points

Crucially, Lift is always shown separately from ratings, and money never touches scores. A 0–100 trust score aggregates security, quality, license, maintenance, and community signals so you can compare skills on evidence rather than marketing.

The Notion AI story shows why this matters. Notion AI shares ChatGPT's known weaknesses — hallucination, knowledge cutoffs, token limits — because it runs on the same underlying models, as independent analysis confirms. Users deserve to know that before they rely on the output. The same logic applies to every skill you attach to an agent: if you can't verify what's inside it, you can't trust what comes out of it.

SkillzBench is in active development right now. If you want AI skills that are proven — verified, scanned, and benchmarked before they touch your agents — join the waitlist at skillzbench.com and be first in line when the platform opens.