I've been using Claude to manage my Notion workspace lately. Creating documents, reorganising content, splitting pages up, moving things around. Claude connects to Notion through an MCP connector, which is basically an API bridge that lets it read and write to my workspace. It works well. Really well, actually.
Better than Notion's own AI.
That should feel paradoxical. Notion has every advantage here. They own the product. They have full context over every page, every database, every workspace. The MCP connector that Claude uses is, by definition, a lossy abstraction over what Notion can see internally. Their AI, all else being equal, should do at least as well as Claude. Probably better.
All else isn't equal, though. And the reason why tells you something important about where the software industry is heading.
The Thousand Percent Gap
When I use Claude on my Max plan to work with Notion, I'm using Opus 4.6. The most capable model available. Generous context. Anthropic has priced my subscription to absorb heavy usage. That's the deal.
The economics of that deal are wild. My Max plan costs a flat monthly fee. The same tokens, purchased through the API at list price, would cost somewhere between ten and a hundred times more. The exact multiple depends on usage patterns, but the directional maths isn't controversial. Subscription users get tokens at a fraction of what API consumers pay.
Now think about what that means for Notion. Their AI feature runs on API pricing. Maybe they've negotiated a volume discount, but they're still far closer to API economics than subscription economics. Every token their AI consumes comes directly out of their margin. Every time a user asks Notion AI to rewrite a paragraph or reorganise a page, Notion is paying for those tokens at something close to retail.
The same work that costs me almost nothing on my Claude subscription costs Notion real money. Per user. Per request.
The Squeeze
This puts SaaS companies in an impossible position. They have two levers to pull, and both make the product worse.
Lever one: use a cheaper model. Instead of Opus 4.6 or something equivalent, they run Haiku, or an open-source model, or some fine-tuned lightweight thing that handles simple tasks but falls over on anything requiring real reasoning. You've probably felt this. The AI in most SaaS tools feels like it's running on a model from two generations ago. That's because it often is. They can't afford the good one.
Lever two: be stingy with context. Feed the model fewer tokens per request. Summarise aggressively. Strip out background information that might help the model make better decisions. This saves money but makes the AI dumber, because a model is only as good as the context it gets. When Claude works with my Notion workspace, it sends generous context. It can afford to. Notion's own AI, paying per token, has to ration.
The result is exactly what you'd expect. The AI features baked into your favourite SaaS tools are, on average, noticeably worse than what you get by pointing a foundation model at the same tool through a connector. Not because the SaaS company has bad engineers. Because the economics won't let them compete.
This Is Dumping
There's a concept in trade economics called dumping. It's when a country sells a product into another market well below cost, with the goal of destroying the local competition. Once the competition is gone, prices come back up. It's considered an act of trade aggression, and most countries have laws against it.
What's happening with AI token pricing has the same structure. Foundation model providers are cross-subsidising usage through subscription plans priced well below API rates. This makes it cheaper for a consumer to use AI through the provider's own interface than through any SaaS product relying on the same models via the API.
Users figure this out quickly. Claude does a better job at tasks inside Notion, Jira, Salesforce, or wherever than those products' own AI features.
The vendor's AI feels like a toy. The foundation model feels like a tool. The pricing structure is a big part of why.
Whether this is deliberate strategy or an emergent consequence of how subscriptions are priced, the competitive effect is identical. SaaS AI features are competing against a subsidised product. That's a fight they can't win on a level playing field.
Find your AI-friendly employer
Browse CompaniesThe SaaSpocalypse Has a Price Tag
People have been predicting the "SaaSpocalypse" for a while. The argument is that AI will collapse the value of many SaaS products by making their core functionality trivially reproducible. Build your own CRM in a weekend, that sort of thing. I've written before about how AI exposes gaps in product thinking. But there's an economic angle here that gets overlooked.
It's not just that AI can replicate what SaaS products do. It's that AI can do it cheaper within the foundation model's ecosystem than the SaaS vendor can do it within their own product. The vendor is paying retail for the same intelligence that the foundation model provider is offering at wholesale. Or below.
This isn't a technology problem. It's a margin problem. And margin problems don't get solved by hiring better engineers or shipping more features. They get solved by finding a different business model. Or they don't get solved at all.
I wrote recently about how token access has become a cultural signal for companies that take their engineers seriously. That same dynamic applies here, but at the industry level. The companies giving their people unconstrained access to foundation models aren't just getting more productive engineers. They're getting better AI than the tools those engineers used to rely on.
Follow the Margin
I'm not predicting the death of all SaaS. Products with deep workflows, proprietary data, and real network effects have moats that survive the pricing squeeze. But the tier of SaaS that's basically "a nice UI over a database with some business logic" is in real trouble. If the AI doing the heavy lifting works better and costs less when accessed through Claude than through the product itself, the value proposition inverts.
The interesting question isn't whether this is happening. It is. The interesting question is whether current subscription pricing is sustainable. Flat-rate token access at these levels is almost certainly a growth-stage strategy, not an equilibrium. At some point, the economics have to balance. But by then, the competitive damage may already be done.
If you want to see this for yourself, try something simple. Pick a task you'd normally ask your SaaS tool's built-in AI to handle, and try the same thing through Claude with an MCP connector. The gap will tell you everything you need to know about where the power actually sits in this new stack.