If you use Notion AI every day, the way you pay for it just changed. On 19 October 2026 Notion split AI usage into two monthly pools, Standard and Select, and made its Auto model router the default way to spend them. Pick a model by habit and you burn the smaller pool first. Leave everything on Auto and you may never notice the limit.
The same update lets Notion agents work on one task for more than 24 hours, up from five minutes. That is a big change for anyone who has tried to hand Notion a research or clean-up job and watched it stop halfway. This guide covers how the credits work, when to override Auto, and how to brief an agent that will run for hours without you watching.
What changed in Notion AI on 19 October 2026?
Notion upgraded its Auto model router, rebuilt the agent harness so a single task can run for more than 24 hours, added a "Review for me" safety mode, let personal agents use a ChatGPT subscription, and replaced its old credit model with monthly Standard and Select credits for Business and Enterprise seats.
The announcement came from co-founder Akshay Kothari in Notion's post "Agents, rebuilt to do more for less". The parts that matter for daily work:
--- Auto router: sends each request to the model that best balances quality, speed and cost. Notion says that across 20+ common use cases it matched Opus 5.5 and Astra at about 90% lower average cost. That is Notion's own test, so treat it as a claim, not a benchmark.
--- Long-running agents: Notion Agent and Custom Agents now run on a durable service that survives restarts, compacts long conversations and offloads large tool results instead of stuffing the context window.
--- Computer and browser use: agents can write and run short scripts, which Notion says is faster and cheaper than doing the same work in model tokens.
--- Review for me: a separate model checks whether an action is safe before the agent takes it.
--- Bring your own AI: connect a ChatGPT Plus, Pro or Go account to your personal Notion Agent (not Custom Agents yet). Bring-your-own-key is in private alpha for Enterprise.
How do Notion's Standard and Select credits work?
Each Business or Enterprise seat gets 1,500 Standard credits a month for work where Notion picks the model, and 500 Select credits for work where you pick the model or run your own Custom Agents. The pools are separate, cannot be combined or transferred, and reset each billing cycle without rolling over.
Notion's help page on managing AI credits gives rough costs: about 5 credits for a simple task, 10 for a medium one and 25 for a complex one. Actual use varies, but the estimates make the trade-off easy to see:
--- 1,500 Standard credits is roughly 300 simple tasks, 150 medium tasks or 60 complex ones a month on Auto.
--- 500 Select credits is roughly 100 simple tasks or only 20 complex runs where you chose the model.
A third bucket sits above both. Workspace "Notion credits", bought by admins, cover high-consumption models such as Fable and Astra, Custom Agents funded by the workspace, and agent runs triggered through the API or a connected app. You can check your own balance under Settings, then Notion AI, then Usage.
Timing matters too. The new system is live for new customers now, and existing customers move over on 2 November 2026. If your workspace renews later, Notion says your account team will set credits at renewal.
When should you let Auto pick the model in Notion?
Leave Auto on for drafting, summarising, translating and editing inside pages, because that work spends Standard credits, the larger pool. Pick a model yourself only when you need one model's specific behaviour, when Auto has failed the same task twice, or when the output goes to a client without editing.
The habit to break is choosing the strongest model for everything. Under the new system that habit moves your everyday work from a 1,500-credit pool into a 500-credit pool, and the premium models may draw on workspace credits instead. Here is a simple rule set you can follow:
--- Auto: meeting-note summaries, rewriting a paragraph, translating a page, tidying a database view, answering a question about your workspace.
--- Pick a model: long reasoning over many documents, a voice-matched draft where you know which model writes closest to your style, or a task Auto keeps getting wrong.
--- Switch back: once the hard step is done, return to Auto for the clean-up. Notion's FAQ suggests exactly this when Select credits run low.
Notion also says Auto will re-route a task to a different model if the first one is not making progress. That makes Auto a reasonable default even for medium tasks, because a stuck run does not quietly eat your credits on one model.
How do you brief a Notion agent that runs for hours?
Give the agent a written finish line, a scope it cannot leave, checkpoints where it records progress, and a clear list of actions that need your approval. A long run without these drifts, repeats work and spends credits on steps you never wanted. Six labelled lines are enough for most jobs.
Five-minute agents forgave vague prompts because they stopped early. A 24-hour agent will keep going, so the brief has to do the supervising you are not there to do. Use this template for a Custom Agent or a long Notion Agent request:
Try this prompt:
GOAL: Build a competitor tracker for our three main rivals in the Hong Kong skincare market.
DONE WHEN: The "Competitors" database has one page per rival with pricing, three latest product launches, and links to every source. Stop when all three pages are complete.
SCOPE: Use only the pages in the "Market Research" teamspace and the public websites of the three brands listed below. Do not edit any other database.
CHECKPOINTS: After each rival, add a line to the "Agent Log" page with what you finished, what you could not find, and the sources you used.
STOP AND ASK: Before deleting anything, sending any message, or changing a page outside the Competitors database, stop and ask me.
IF UNSURE: Mark the field [NOT FOUND] and move on. Never guess a price or a date.
Two lines do most of the work. DONE WHEN gives the agent a reason to stop, which protects your credits. CHECKPOINTS gives you a log to read the next morning, so you can see where it went wrong without replaying the whole run.
Pair this brief with the new "Review for me" approval mode for any agent that can write to shared pages. The checker model is a safety net, not a substitute for the STOP AND ASK line.
What are the limits and gotchas of the new Notion AI?
The 90% saving and the 10x credit claim come from Notion's own tests. Credit costs per task are estimates. The ChatGPT connection works only in your personal Notion Agent, the AI Command Center is announced as coming soon, and existing customers keep their old terms until 2 November 2026.
Before you rebuild your workflow around the update, check these:
--- The benchmarks are self-reported. Notion's chart compares quality against credits on its own task set. Run your three most common jobs on Auto and on your favourite model, then compare the outputs yourself.
--- Long runs are not free runs. A 24-hour task can still burn a month of Select credits if it loops. Start with a short DONE WHEN and widen it once you trust the agent.
--- Team Custom Agents spend Select credits by default, even when colleagues trigger them. Ask your admin whether a shared agent should be funded from workspace credits instead.
--- Plan scope. The Standard and Select allowances are described for Business and Enterprise seats. If you are on another plan, check the pricing page for what applies to you.
--- ChatGPT connection limits. Using your ChatGPT plan inside Notion keeps your Notion credits untouched, but it is personal-agent only for now and your ChatGPT plan limits still apply.
If you already pay per step for AI in an automation tool, the same discipline applies. Our guide to AI step costs in Zapier walks through a similar audit.
How can you test Notion's new credits in 20 minutes?
Run one week's worth of your usual Notion AI jobs on Auto, check the Usage page, then run one long task with the six-line brief. Compare credits spent with the 1,500 and 500 allowances. You will know within 20 minutes whether your habits fit the new pools or need changing.
--- Minutes 0 to 5: open Settings, Notion AI, Usage and note your current Standard and Select balances.
--- Minutes 5 to 12: on Auto, run five tasks you do every week, such as summarising a meeting page or rewriting a brief. Note the credits used.
--- Minutes 12 to 20: start a small agent task with the six-line brief, scoped to one database. Check the Agent Log and the Select balance when it finishes.
Multiply the five-task total by four for a monthly estimate. If it is well under 1,500, your daily work is safe on Auto and you can save Select credits for the agents that matter. If you are building agents in other tools too, our n8n agents first-build guide uses a similar instruction brief. The tools keep changing; clear briefs and a credit check keep you in control. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team. Details reflect Notion's 19 October 2026 announcement and its "Manage your AI credits" help page as read on 22 October 2026. Credit allowances, plan coverage and rollout dates may change.
Give your agents ready-made skills
A well-briefed agent is only half the job; it also needs the right skills for the role. Browse UD's free AI Employee Hub for role-based skills you can download today, and when you want them running in your daily workflow, we'll walk you through every step, from tool setup to workflow design and deployment.