The most interesting line in Notion’s latest AI update isn’t about a smarter agent. It’s about the bill. Notion says its upgraded model router, Auto, delivers results close to frontier models at roughly 90% lower average cost. If that holds up, it changes how teams think about running agents all day.
Notion announced the changes on October 19, and there are several pieces. The company rebuilt its agent harness (it says this is the second time), and Notion Agents now get their own computer and browser, better context management, and a new “Review for me” mode with safety checks. They can work on a task for up to 24 hours, and chats load up to 50% faster on mobile.
Auto is the headline. It works as a router: it matches each task to a model based on quality, speed, and cost, and changes course if the first attempt stalls. Notion says that across 20+ common tasks, Auto produced results similar to Opus 5.5 and Astra at about 90% lower average cost. In practice, credits go about 10 times further. Those are Notion’s own numbers, and the task set isn’t public, so treat them as a claim to test on your own workflows. Still, the direction is right. Most users don’t want to pick between models, and routers from OpenAI and others are chasing the same idea.
The more interesting strategic move is bring-your-own-AI. Users can connect a ChatGPT Plus, Pro, or Go plan to their personal Notion Agent. Bring Your Own Key is also in private alpha for Enterprise customers. That’s a notable concession. Notion is saying it doesn’t need to resell every token, and it would rather be the place where work happens regardless of who supplies the model. For people already paying OpenAI, that avoids paying twice.
Enterprise buyers get something too. The AI Command Center, coming soon, will show usage by person, feature, and agent. Admins can set spending limits and alerts and decide which models and agents their team can use. Anyone who has watched agent costs creep up across a company will see why this matters. Governance is becoming a requirement, not an extra, and Microsoft and Google are building similar controls into their own suites.
Then there’s MCP. Notion calls its Model Context Protocol support its fastest-growing product ever, and it has overhauled it so outside agents can treat Notion as a shared brain: find context, remember what matters, and create shareable output. Memory sync is live today, with skills via MCP coming soon. This puts Notion in the same conversation as Claude, ChatGPT, and Cursor users who want one workspace behind their agents, rather than competing with them head on.
All of this runs on a new credit system. Notion Agent, Custom Agents, and many outside AI tools will draw from a single pool. That’s simpler than separate meters, but credit systems often hide real price changes. Watch how many credits common tasks actually consume, and what each plan includes.
Does it matter? Mostly yes. Longer-running agents, cheaper routing, and cost controls are the unglamorous features that decide whether companies keep paying for agents after the pilot. The 90% figure needs independent testing. But the combination of lower cost, outside model access, and admin oversight is a coherent pitch, and it’s aimed squarely at the budget conversation.
