Notion agents can now work for 24 hours. Here's how to use them without burning credits
Until last week, a Notion agent could work on a single task for about five minutes. That was enough to polish an email or summarize a meeting, but not enough to clean up a database, work through a backlog or research a list of prospects. Anything bigger, you either split into pieces yourself or took somewhere else.
On October 19, Notion raised that limit to more than 24 hours. It also gave agents their own computer and browser, made its model router much cheaper, and changed how AI usage is paid for. The longer tasks get the headlines, but the credit changes are what will decide whether this saves you money or quietly eats it.
What actually changed
Notion rebuilt the system that runs Notion Agent and Custom Agents. It is the sixth version, and the main changes are these:
- Long-running tasks. Each agent session now runs on its own service and keeps going through restarts, which is how a single task can run for more than 24 hours.
- Computer and browser use. Agents can write and run short scripts instead of doing everything through the language model, which Notion says is faster, cheaper and more accurate. They can also work in a browser.
- A safer approval mode. A new "Review for me" setting uses a separate model to check whether an action is safe before the agent takes it.
- Lower running costs. Notion says Custom Agent runs cost about 30% less in testing, and agents run up to 50% faster on web and mobile.
- More tools. Agents can now reach hundreds of tools through MCP and connectors, up from a few dozen earlier this summer.
Leave it on Auto
The single biggest lever on your credits is the model setting.
Auto is Notion's model router. It reads each task, sends simple work to fast, cheap models and harder work to more capable ones, and switches models if one gets stuck. Across more than 20 common tasks, Notion says Auto matched the results of top models like Opus 5.5 and Astra at about 90% lower average cost. Those are Notion's own tests, so treat the exact figure with some caution, but the gap in its own credit estimates is large:
| Task | Example | Auto | Opus 5.5 picked by hand |
|---|---|---|---|
| Simple | Polish an email, translate a document | 5 credits | 50 credits |
| Medium | Write a work report, triage support inquiries | 10 credits | 85 credits |
| Complex | Maintain CRM contacts, configure databases | 25 credits | 150 credits |
The habit to build is simple. Leave Notion Agent on Auto for everyday work, and only pick a model yourself when Auto's output genuinely isn't good enough for that specific job.
How the new credits work
If you're on a Business or Enterprise plan, each seat now gets two separate monthly allowances:
- 1,500 Standard credits for work where Notion picks the model, which means anything on Auto.
- 500 Select credits for work where you pick the model yourself, and for Custom Agents you set up.
The two pools are separate. You can't combine them, move them between people, or carry unused credits into next month. When one runs out, the work that uses it pauses until your monthly reset, unless your admin has turned on extra workspace credits.
Some usage skips both pools and bills straight to the workspace's paid Notion credits. That includes the most expensive models like Fable and Astra, Custom Agents funded by the workspace, and agents run through the Notion API or a connected app.
Two dates matter. New workspaces moved to this system on October 19, and existing workspaces move on November 2. From December 1, AI Meeting Notes also starts using Standard credits, at roughly 10 credits for an hour-long meeting.
Watch who pays for shared agents
This is the detail most teams will miss. A Custom Agent you build runs on your Select credits, even when you share it with your team and other people use it. One popular agent can drain its owner's 500 credits in the first week of the month.
If an agent is doing work for the whole team, ask your admin to fund it from the workspace's Notion credits instead. That keeps your own allowance for your own work, and it gives the admin one place to see what the agent actually costs.
What to hand a 24-hour agent
The longer limit changes which jobs are worth giving to an agent. A few that now fit in a single task:
- Database cleanup. Merge duplicates, fill in missing fields and standardize tags across a CRM or content database that has drifted for months.
- Backlog processing. Work through a quarter of meeting notes and pull out every decision and open action into your task database.
- Research with the browser. Take a list of companies, look each one up, and fill in a database row for each with what they do and who to contact.
- Recurring reports. A Custom Agent that compiles your weekly numbers from several sources into one page every Monday.
Before you set a long task running, test it on a small batch first, such as ten rows instead of five hundred. A long task that goes wrong also goes wrong for a long time, and it spends credits the whole way. Turning on "Review for me" for anything that edits or sends things is a sensible default.
Use the AI you already pay for
You no longer have to pay for AI twice. If you have a ChatGPT Plus, Pro or Go subscription, you can connect it to your personal Notion Agent and use it instead of your Notion credits. It applies when you pick a GPT model yourself, and it doesn't cover Auto or Custom Agents yet. Enterprise customers can also try bringing their own model key, which is in private alpha.
Notion also overhauled its MCP, which lets outside agents like ChatGPT and Claude use your Notion workspace as shared context. If you already do most of your thinking in another AI tool, this is the route to connecting it to Notion rather than moving everything across.
What to do before November 2
If you already have a Notion workspace, spend twenty minutes on this before the switch:
- Check your usage. Go to Settings, then Notion AI, then Usage, and see how close your current habits are to 1,500 Standard and 500 Select credits.
- Switch your agent to Auto. If you've been defaulting to a top model by hand, this is the change that saves the most.
- Audit your Custom Agents. List who owns each one and who uses it, and move any team-wide agents onto workspace funding.
- Decide on overflow. Admins can turn on "Allow workspace to use Notion credits after AI limit is reached" in Settings, then Notion AI. Turn it on if a pause mid-month would hurt you, and leave it off if you'd rather have a hard spending cap.
- Pick one long task to test. Choose the job you've been putting off because it was too big for a five-minute agent, and run it on a small batch.
The 24-hour limit is the exciting part. Getting the credits right is what makes it worth using.
Related guides
GPT-6.1 Sol vs Claude Sonnet 5.5: which model for your builds
Two cheap frontier models landed within a week. Here is where each one wins, and how to switch without breaking your automations.
Jev explained simply: when to use it instead of ChatGPT
A plain-English guide to Jev, TypeSafe's new decision model: what it does, the jobs it handles better than an LLM, real examples from the first two weeks, and how to add it to a no-code workflow.
Stop using humans to run your Google Ads
A practical eight-step guide to setting up Betatron, the AI agent that builds and runs Google Ads campaigns inside your own account.