Skip to main content
Bargain brainpower

Notion's Auto router cuts AI costs about 90%

Notion is rethinking how its AI spends credits, with a new model router that picks the right model for each task and a rebuilt agent system that can now keep working for much longer. Here's what changes for users, admins, and anyone already paying for AI.
Notion's Auto router cuts AI costs about 90%
Notion's Auto router cuts AI costs about 90%

Notion is overhauling how its AI spends credits. Its new model router, Auto, delivered results similar to Opus 5.5 and Astra across more than 20 common use cases at about 90% lower average cost, so credits go about 10× further. The update also includes a rebuilt Notion Agent harness, a new credit system, and new ways to use AI subscriptions you already pay for.

Notion says AI is moving fast. In the past month it added seven new models with partners Anthropic, OpenAI, and open-weight providers. Notion Agents can now work on a single task for more than 24 hours, up from five minutes. Agents can also connect to hundreds of tools through MCPs and connectors, up from a few dozen earlier this summer. Notion describes its goal as offering “the best model-agnostic context layer” for use with any model from any provider, since customers increasingly use Notion as the infrastructure for their own agents.

Auto: the model router that picks for you

Choosing among many models is hard, and people often default to expensive models for most requests. Auto reads each task and sends it to the model that best balances quality, speed, and cost.

Auto handles simple requests with fast, efficient models. Longer or harder work, such as analysis or multi-step agent runs, goes to more capable models. If a model is not making meaningful progress on a task, Auto re-routes the work to a different model.

New models arrive almost every week. Notion evaluates them, and Auto can begin using providers that an admin has approved without users needing to track each release. Notion says it constantly tunes Auto on real knowledge work, with one goal: “the best, fastest result for the least credit consumption.”

A sixth-generation Notion Agent harness

Notion Agent
Notion Agent — Notion

Notion rebuilt the Notion Agent harness again, and this is its sixth generation. Shir, Notion's Head of AI and Core Product, explained the aim in a team meeting:

To support long-running threads in a way that wouldn't blow through credits, we knew we needed to pair them with significant efficiency gains.

— Shir, Notion's Head of AI and Core Product

Several changes make that possible:

  • Durable execution for long tasks. Every session runs on a dedicated service with a queue-backed worker loop, so work continues through restarts and hiccups. This is how a task can now run for 24 hours.
  • Computer and browser use. Agents have their own computer and can write short scripts. Notion says these scripts are faster, cheaper, and higher quality to run as code than to execute with language model tokens.
  • Smarter context management. Agents compact long conversations, using each model provider's native compaction where possible. Large tool results are offloaded instead of filling the context window. Notion says that means less to reread on every step and less compute per task.
  • Faster apps. Agents run up to 50% faster on both web and mobile.

The result, Notion says, is agents that finish more of what they start. Its infrastructure uptime is now above 99.9%, and Custom Agent runs cost about 30% less in testing.

Use the AI you already pay for

Notion says you shouldn't have to pay for AI twice, so it added several ways to use existing AI access:

  • ChatGPT subscriptions. In close partnership with OpenAI, Notion built a connection that lets you use a ChatGPT Plus, Pro, or Go account inside Notion. It works in your personal Notion Agent to start, not in Custom Agents. More plans and surfaces will follow.
  • Bring Your Own Key (BYOK). BYOK is in alpha for Enterprise plan customers, letting them use their own model-provider key. Notion points to an alpha onboarding guide and says customers should talk to their account team about access.
  • An overhauled Notion MCP. The MCP connects agents such as ChatGPT, Claude, Grokbot, and Muse to Notion as a shared brain, where they can find context, remember what matters, and build shareable artifacts. Notion calls MCP its fastest-growing product in company history, so it invested more. Users can now sync memories to Notion, and Skills via MCP are coming soon.

Admin visibility with the AI Command Center

Notion AI command center
Notion AI command center — Notion

Admins need to see what is running, who owns what, and what it costs. Notion is building an AI Command Center that brings these views into one place, which it says unifies controls previously spread across multiple locations.

Admins will see credit usage by member, feature, and agent, with alerts as the workspace nears its limit. They can set spending limits for members and agents, choose which models members can use, and control who can create, pause, or archive Custom Agents. Members can see how many of their included credits they have used and what each of their agents spends.

A new credit system

Business and Enterprise seats come with two monthly credit types:

  • 1,500 Standard credits per month for work where Notion picks the model.
  • 500 Select credits per month for work where you pick the model, or for running your Custom Agents.

Admins can buy additional credits for their workspace and control how they are allocated. Additional credits can cover Custom Agents shared with the workspace, high-credit models (for example, Fable and Astra), and any usage beyond included credits. Users can also try Custom Agents without paying first, using their included monthly Select credits.

Notion says the upgraded router means credits go much further, and based on its estimates, the vast majority of users are unlikely to run out. The two pools stay separate because AI tasks consume credits at very different rates. A few complex tasks or frequently running Custom Agents could quickly exhaust a shared allowance. Separate pools protect everyday usage on Auto while still letting users choose models and run agents.

Availability

  • New customers: the new system is live today.
  • Existing customers: they move over on November 2, and admins will receive an email with details.
  • Contracts not yet renewed: customers move to the new system on November 2 with the same Standard and Select credits as everyone else. At renewal, credits are set by plan, and the account team will walk through the details.
  • AI Command Center: coming soon.
  • BYOK: alpha for Enterprise plan customers, by request through their account team.
  • ChatGPT subscription connection: available for personal Notion Agents, with more plans and surfaces to follow.
Feature Previous New
Model selection One model for the task Model selected per task step
Model switching Did not switch during a task Re-routes if progress stalls
Optimization Not specified Balances quality, speed, and cost
Agent runtime 5 minutes More than 24 hours
Average cost Not specified About 90% lower

Felipe Santos

“Artificial intelligence can process the world in milliseconds, but only the human heart can give meaning to every second lived” – Mr. Santos