Notion's Credit-Based AI Pricing and the Trust Layer for Agents

Notion rebuilt its AI pricing around a credit system while publishing exactly how its model router actually works, 1Password shipped a trust layer specifically for Meta's Muse agent, and Dare Obasanjo's read on Meta's earnings call numbers gives us the clearest data point yet on how fast people adopt an agent when it ships inside apps they already use. ### Notion - Rebuilt AI Pricing Around a Credit System, Then Showed Its Work on Model Quality **Source:** https://www.notion.com/blog/notion-ai-more-for-less **Credibility:** High (first-party product announcement, paired with a companion engineering post explaining methodology) **What happened:** Notion rebuilt Notion AI around smarter model routing, long-running agents, bring-your-own-AI support, and a new credit-based pricing system. Akshay Kothari, Notion's co-founder, framed it as "more for less": faster responses and lower cost per task, in exchange for moving away from flat unlimited-feeling AI access toward a credit meter. Alongside the pricing post, Notion published a second piece from engineering manager Ghita Houir Alami explaining how the upgraded model router actually gets evaluated: pairwise comparisons and calibrated judging used to measure model quality at scale before routing decisions get made. **Key technical details:** - New credit system replaces (or supplements) previous AI access model, tied to usage rather than flat subscription inclusion - Model router now picks between models per-task using calibrated judging, not a static default - Bring-your-own-AI option lets teams plug in their own model/API keys instead of using Notion's default routing - Long-running agents are now a supported product surface, not just single-turn Q&A - Companion post details the actual evaluation methodology (pairwise comparison + calibration) used to decide which model handles which task **Why it matters for PMs:** This is the second time in two weeks Notion has published not just what they shipped but how they decided it was good enough to ship (see also their 160-apps-in-three-weeks agent post). That's rare and worth studying: most companies say "we use the best model for the job," Notion is showing the actual pairwise-judging pipeline behind that claim. The credit-system move also answers one of our open questions directly: Notion is betting users will tolerate metered AI pricing if it's paired with visible quality and speed gains, rather than assuming AI has to feel free and unlimited to be adopted. **Critical questions:** - Will users perceive credits as a downgrade even if the underlying model routing genuinely improves task completion speed? - How transparent is the credit consumption to end users in the moment, before vs. after a task runs? - Does bring-your-own-AI cannibalize Notion's own credit revenue from power users who'll just plug in their own API key? **Action you could take today:** If your product has flat-rate AI access, pull your cost-per-task data and model it against a credit system like Notion's, specifically where your heaviest users sit relative to a hypothetical credit cap.