Notion moves AI agents to credit pricing and pitches itself as the context layer

In an interview published on 28 October by the French outlet Blog du Modérateur, David Tibbitts, product marketing at Notion, explains two shifts at once. Notion now presents itself as a "context layer" for AI agents, a neutral "Switzerland of AI" rather than only a connected workspace. And it has moved away from flat AI pricing towards metered credits. The second shift is the more instructive one for anyone selling software with AI inside.

From flat AI to metered credits

The timeline given in the article is short. In February 2023, Notion AI became a paid add-on billed per member. In May 2025, AI was included in the Business plan for new subscribers. In May 2026, Custom Agents moved to credits, at $10 per 1,000 credits. In October 2026, Workers, the developer platform, followed.

The reason is cost. According to Tibbitts, a Notion agent could run autonomously for five minutes at most twelve months ago; it can now run for up to 24 hours. An all-you-can-eat plan cannot absorb that range of usage, and Notion says it refuses to subsidise tokens at a loss, which it says many AI companies do. Seats remain for the software itself; AI is sold as consumption on top.

Two other elements matter. An "auto" mode routes each request to the model judged most suitable, balancing quality, speed and cost, while users can still pick a model themselves. And administrators get controls: visibility on spending, the ability to restrict who can create agents, and the option to allow or block expensive models. Tibbitts also mentions outcome-based billing as an interesting idea, without announcing it.

What the interview does not say is just as relevant: how many credits an action or a model consumes, how many credits come with the Business plan, and what the 2023 add-on cost. Without these figures, the real price for a given team cannot be calculated from the article.

The context layer as a retention argument

The positioning side is a retention play. Notion works with OpenAI and Anthropic and with open-weight models it does not name, and says new models often become available in Notion within minutes of release; that is the company's claim, not something the article verifies. If models are interchangeable and agents are a crowded market, which Tibbitts acknowledges, the durable asset is what agents need to work: company context, skills, memory, permissions and governance. Notion describes this as a backpack the user carries from one agent to another.

For a SaaS company, this is the logic of switching costs moved one level down. The model is a commodity; the accumulated, permissioned, shared context is not. The argument also leans on Notion's collaborative origins: results land on a shared surface, and agent settings can be adjusted by teammates rather than living in one person's chat history.

What to check in your own pricing this week

If you sell software with AI features, the Notion case raises questions you can answer with your own data:

• Measure the distribution of AI cost per account, not the average; long-running or agentic features stretch the tail, and the tail is what breaks a flat price.
• Identify which AI features are short, predictable tasks that fit inside a seat price, and which are open-ended work that should be metered.
• If you introduce credits, publish what an action costs in credits; buyers who cannot estimate a bill tend to delay or cap usage.
• Give administrators spending visibility and model restrictions before they ask; in team accounts, the person who approves the budget is rarely the person who triggers the usage.
• Decide what context your product holds that a competitor's agent would need, and make sure customers can see and use it, since that is what keeps them when models change.

The broader signal is that flat AI pricing is being abandoned by a company that once bundled AI into its main plan. Any team still offering unlimited AI should know exactly which customers it is subsidising, and why.