Notion's router matched frontier models at about 90% lower cost. Most agent steps don't need the big model.

Notion said on October 19 that its Auto router matched Claude Opus 5.5 and GPT-6 Astra at about 90% lower average cost, which tells me most agent steps never needed the big model. In Notion's October 19 post, co-founder Akshay Kothari says the test covered 20+ common use cases. Auto sends simple requests to fast, cheap models. Longer work, like analysis or multi-step agent runs, goes to stronger ones. If a model stalls, Auto re-routes.

The same post says Notion agents can now work on one task for more than 24 hours, up from five minutes. Long runs make routing matter more. A day of steps priced at the top model adds up.

Here's what I'd do with my own agent. Tag each step by the judgment it needs. Fetching a page is cheap. So is filling a field or summarizing a thread. Deciding whether to act is the expensive step. Pay for a strong model there, and only there. Then read the bill a week later.