“Once you use up all Standard or Select credits, impacted work will pause until your credits reset.” Notion Help Center, “Manage your AI credits”, read October 20, 2026.
That line is on Notion’s help page, not in the launch post. Notion rebuilt its agents on Monday, October 19. One task can now run for more than a day, up from five minutes, and most of the coverage stopped at that sentence. If you’re on a Business or Enterprise plan, leave the model picker on Auto. Don’t pick a model yourself unless Auto got it wrong, because the pool for hand-picked models is a third the size and drains about ten times faster. And ask your admin to switch on the overflow setting before November 2, or a support reply, an inbox clean-up or a follow-up list can stop halfway through the month.
What nobody’s written up is what a credit buys. Notion’s help page carries a table of its own estimates, so I divided a seat’s monthly pool by each job’s price. The jobs-a-month columns are mine, rounded down to whole jobs. The credit prices are Notion’s, from its own replay tests, which is one source.
| Job, in Notion’s own examples | Auto, credits per job | Jobs a month from 1,500 Standard credits | Opus 5.5, credits per job | Jobs a month from 500 Select credits |
|---|---|---|---|---|
| Simple: polish an email, draft a support reply, translate a message | 5 | 300 (1,500 / 5) | 50 | 10 (500 / 50) |
| Medium: write a work report, triage support, manage your inbox, summarise an hour-long meeting | 10 | 150 (1,500 / 10) | 85 | 5 (500 / 85 = 5.9) |
| Complex: run hiring, set up a database, refine a quarterly plan, keep CRM follow-ups current | 25 | 60 (1,500 / 25) | 150 | 3 (500 / 150 = 3.3) |
Credit prices from Notion’s credits page, read October 20, 2026. Division done here, rounded down.
So a seat that leaves Auto on gets somewhere between sixty big jobs and three hundred small ones a month. A seat that picks Opus 5.5 by hand gets three big jobs, or ten email polishes, and then it waits. The two pools don’t mix. Running out of one doesn’t touch the other, and leftover credits don’t roll over. How many of your team’s jobs are the complex kind? That’s the number that decides whether a seat lasts the month.
What Changes on November 2
New workspaces are on the credit system already, as of October 19. Existing ones move over on November 2, which Notion states on the help page and the launch post. Three things to do before then:
- Open the credits dashboard in Settings. Each member shows as a row, so you can see who’s spending what before the clock resets.
- Find the switch called “Allow workspace to use Notion credits after AI limit is reached”. It’s off by default. Off means a member who runs dry waits for the reset. On means the work carries on and the workspace pays.
- Leave Auto on. Hand-pick a model only when Auto missed, then switch back.
One more line worth knowing. From December 1, meeting notes start drawing on the Standard pool, at about ten credits for an hour of meeting. How many hours of meetings does your team sit through in a month? A seat in a lot of meetings gets a smaller pool than the headline says.
The launch post says Auto matches frontier models at about ninety percent lower cost. I checked that claim against Notion’s own pages. The blog says ninety, the Auto help page says eighty-seven, and both come from Notion replaying its own tasks in a test copy of Notion. That’s Notion marking its own work, and nobody outside has checked it yet. I haven’t run the new build myself.
What Notion won’t tell you is which model Auto picked. The help page says there’s no published list and you can’t see which model handled your task or how often. For a support reply that probably doesn’t matter. For anything with a client’s numbers in it, ask Notion before November 2 which providers your workspace allows, because that’s the one lever you do have.
The Same Day, Reuters Found Agents That Need a Day to Break In
Reuters reported on Monday that one Chinese-speaking attacker used an open-source agent platform called Cairn to break into more than a hundred company websites, mostly online shops in the United States, Canada, Australia and Japan, for about five thousand dollars in model usage. The agents tried over 1,200 sites since late September and got into roughly one in ten. The security firm that found the control panels, NorForge, put the cost at around fifty dollars per successful break-in. A successful one took the agents a little under 23 hours on average.
The same day Notion sold a day-long agent as a feature, Reuters found the other side already had one. Reuters’ page turned my fetch away, so the figures here come from The Next Web and CyberSecurityNews, both quoting the Reuters piece.
The brand names got the headlines. Deere said it had fixed the hole and had no sign anyone reached sensitive data. SharkNinja said the hit was a standalone product support site, now contained. Cummins said it was looking into a vendor’s site. What none of the coverage says plainly is what the holes were, and it’s the most useful part. NorForge’s list, as carried by CyberSecurityNews: database commands typed into web forms, file uploads that nobody checked, login screens that could be stepped around, an old component of the Zend framework, and a session store left open for writing. Those are old web-shop flaws. The agents didn’t invent anything. They were patient, and they were cheap.
What they left behind is an uploader form, a page planted on the server so the attacker can drop files later. The fix for that class of hole is a software update and a closed upload form, not an AI product. If your store runs on software you installed before 2026, ask whoever looks after it two questions this week. What version is it on? Can anyone upload a file to it without logging in? If the answer to the second one is a shrug, that’s the job for this week. The agents needed a day per site. Your update takes less.
Two things aren’t known. Reuters didn’t name which shop platforms were hit, so I can’t tell you whether yours is on the list. And the worker name in the attacker’s panel suggested a DeepSeek model driving a Claude Code setup, which NorForge flagged and nobody has confirmed.
If you’d rather see where your hours go first, the free time audit is six questions, about ninety seconds. Skip it if you already know.