Notion Just Made Picking Your Own AI Model the Premium Option

Notion rebuilt its AI pricing on Monday, and the most interesting detail is not the headline about agents that can now work on a single task for more than 24 hours. It is how the new credits are split. Business and Enterprise seats now get 1,500 Standard credits a month for work where Notion picks the model, and 500 Select credits a month for work where you pick the model or run your own custom agents. Choosing a specific model is now the scarcer, separately budgeted option.

That is a small line in a pricing page, but I think it says a lot about where the AI market is heading. The company that holds your work is starting to decide which model does the work. The model labs are becoming suppliers behind that decision.

The router is the product now

The engine behind this is Auto, Notion's upgraded model router. According to Notion, Auto reads each task and sends it to whichever model best balances quality, speed, and cost. Simple requests go to cheap, fast models. Harder analysis and multistep agent work go to more capable ones. If a model stalls, Auto switches to another. Notion says that across more than 20 common tasks, Auto delivered results similar to the top frontier models at about 90 percent lower average cost, so credits go roughly ten times further. That is Notion's own testing, not an independent benchmark, but the direction is believable.

Look at the incentive this creates. Notion said it added seven new models from Anthropic, OpenAI, and open weight providers in the past month alone. When a cheaper model gets good enough for summarizing meeting notes or drafting a project update, Auto can quietly start using it. The customer does not notice. Notion's margin does. Every week that open models close the gap on ordinary work, the router gets to route more of that work away from the expensive option.

This is exactly the frontier premium collapse I keep talking about, except now it is showing up inside a mainstream business app rather than on a hobbyist's workstation. Most people were never paying for the best model because they needed it. They were paying for it because picking a model is confusing and the expensive one feels safe. Notion is betting it can absorb that confusion and keep part of the savings.

Who gains leverage

The obvious winner is Notion. It already holds the documents, the meeting notes, the project databases, and the company wiki. That context is what makes an agent useful. The model is the part you can swap. Notion even describes its goal as offering the best model agnostic context layer for any model from any provider. When the context owner also controls routing, it becomes the buyer of intelligence on behalf of millions of seats, and buyers with that kind of volume get to negotiate.

The model labs get a more complicated deal. They gain distribution, because their models show up inside a tool people already use every day. But they lose the direct relationship. A user on Auto does not know or care which lab answered the question. The brand that matters is Notion. If a lab wants to be chosen, it has to win on Notion's internal evaluations, not on mindshare. That is a much tougher position than selling a famous chatbot subscription directly to a person.

There is one interesting exception. Notion also built a connection so people can use their ChatGPT Plus, Pro, or Go subscription inside their personal Notion Agent, and it opened bring your own key in alpha for Enterprise customers. I read that as OpenAI choosing to stay visible inside someone else's product rather than become an invisible supplier. Expect other labs to push for the same arrangement. Being a named option is worth a lot when the default is a router that never says your name.

The lock in moves, it does not disappear

I have argued for a long time that building model agnostic is the right defense against vendor lock in. Notion's move is a reminder that model agnostic for whom matters. Your workspace is now model agnostic. You are not necessarily freer. Your dependence just shifted from a model lab to the app that owns your context and decides the routing.

The two credit pools make that dependence concrete. Work on Auto is cheap and abundant. Work where you insist on a specific model, or run your own agents, draws from the smaller pool. Notion's own explanation is reasonable: complex tasks and frequently running agents can drain credits fast, so separating the pools protects everyday usage. But the effect is still that the default path, where Notion decides, is the one priced for heavy use. Defaults shape behavior, and over time most teams will let the router decide.

Notion's improved MCP connection, which it calls its fastest growing product ever, points the same way. It lets outside agents like ChatGPT and Claude use Notion as a shared brain, with memory sync live now. That is generous on the surface. It also means Notion wants to be the place where your memory lives no matter which agent you use. If your company's accumulated knowledge sits there, Notion holds the most durable asset in the stack.

What this means for owning the stack

The structure itself is familiar to me. I already split my own AI work between interactive models and background agents running on hardware I own, because not every task deserves the same model. Notion has turned a version of that idea into a pricing model for a mass market. The difference is who owns the routing decision and who keeps the savings.

For a founder, I think the honest read is that the margin in AI is moving to whoever controls three things: the context, the routing, and the default. Model quality still matters at the very top end, and some tasks will keep justifying the expensive model. But for the bulk of knowledge work, the competition is turning into a contest over who sits between the user and the models. Notion just made a strong claim to that seat for a lot of offices.

My prediction is that more SaaS companies will copy this structure within the next year: a cheap default lane where the vendor picks the model, and a pricier lane where the customer does. The test I would apply to any AI product your company depends on is simple. Ask who decides which model runs your work, who keeps the savings when that model gets cheaper, and whether you could move your context somewhere else if the answer stopped working for you.