The biggest thing in Notion’s October 19 Notion AI release is not the 24-hour agents. It’s the money. Notion rebuilt its agent harness, upgraded the Auto model router, and swapped in a two-tier credit system that hits existing workspaces on November 2 [3][4]. Everything else — BYOK, the MCP overhaul, the AI Command Center — is real but narrower, and one of those items isn’t shipping yet.
So here’s the frame I’d use as an admin. Capability went up modestly. Cost per task is the thing Notion is actually promising to cut. Trust the second claim less than the marketing suggests, and test it before you migrate a whole workspace onto it.
What actually shipped in the October 19 Notion AI release
Six items, roughly in order of how much they’ll change your month:
- Rebuilt agents. Notion Agent and Custom Agents run on a new harness — Notion says it’s the sixth generation — with longer task runs, their own computer and browser, tighter context handling, and a safety-check mode [1][3].
- Auto, upgraded. The model router now picks a model per task and re-routes when a model stalls [3][4].
- Bring your own AI. ChatGPT Plus, Pro, and Go plans can connect to a personal Notion Agent, and Bring Your Own Key (BYOK — using your own provider API key) is in private alpha for Enterprise [3][4].
- Notion MCP. The server that lets outside agents read and write your workspace got an overhaul, with memory sync live now [3][4].
- AI Command Center. Admin dashboards and spend controls. Previewed, not available [1][3][4].
- New credits. Standard and Select pools, monthly, included with Business and Enterprise seats [3][4].
The rebuilt agent harness: 24-hour tasks without blowing credits
Agents can now work a single task for more than 24 hours, up from five minutes [3]. That’s the headline number, and it’s a genuine architectural change rather than a prompt tweak.
How durable execution keeps agents alive
Each session runs on a dedicated service with a queue-backed worker loop, which is how work survives restarts and hiccups instead of dying mid-task [3][4]. Notion says infrastructure uptime is now above 99.9% [3][4].
Two efficiency tricks sit alongside it. Agents get their own computer and browser, so they can write short scripts and run them as code instead of burning tokens on the same work [1][3]. And long conversations get compacted, using each provider’s native compaction where possible, while large tool results are offloaded rather than dumped into the context window [3][4]. Less re-reading per step means less compute per task. That’s the mechanism behind the cost claims, and it’s the part I find most credible.
Safety, speed, and the 30% cheaper claim
A new “Review for me” mode uses a separate model to judge whether an action is safe before the agent takes it [3][4]. Sensible. If you’re letting an agent touch databases unattended, you want a second opinion in the loop.
Notion also says agents run up to 50% faster on web and mobile, and that Custom Agent runs cost about 30% less in its own testing [3][4]. Caveat: those are vendor numbers. No third party has published a replication, and Notion hasn’t detailed how credits map to individual tasks or models. Treat 30% as a direction, not a budget line.
Auto model routing and the 90% cheaper claim
Auto reads a task and sends it to whichever model balances quality, speed, and cost — cheap models for simple requests, stronger ones for analysis and multi-step agent runs [3][4]. If a model stops making meaningful progress, Auto re-routes [3][4].
The number to scrutinize: across 20+ common tasks, Notion says Auto matched Opus 5.5 and Astra at roughly 90% lower average cost, which it says makes credits go about ten times further [1][3][4]. That is a big claim, and it comes entirely from Notion’s own testing [4]. No independent benchmark backs it up yet.
My read: the mechanism is sound. Routing routine work to smaller models is exactly how you cut spend, and re-routing on stalls fixes the usual failure mode of cheap-first routers. Whether the real figure is 90% or 50% matters less than the pattern — if Auto is anywhere near right, defaulting everyday work to it is free money. Just measure your own before-and-after credit consumption rather than assuming.
Bring your own AI: ChatGPT plans and BYOK
You can now connect a ChatGPT Plus, Pro, or Go subscription to your personal Notion Agent, built with OpenAI [3][4]. One catch that will trip up team leads: it works in personal agents only, not Custom Agents, with more surfaces promised later [3][4].
BYOK is in private alpha for Enterprise plan customers [1][3][4]. Notion says to talk to your account team about access; the alpha intake process isn’t documented publicly. If you’re not on Enterprise, this isn’t a decision you can make yet.
Notion MCP: shared brain for outside agents
Notion calls its MCP server the fastest-growing product in company history [1][3][4]. Whether or not that framing holds up, the overhaul is aimed at a real workflow: outside agents like ChatGPT, Claude, and others using Notion as durable memory rather than a read-only document dump. Memory sync is live now, and skills over MCP are described as coming soon [3][4].
Notion AI credits: Standard vs. Select, and the November 2 migration
Business and Enterprise seats include 1,500 Standard credits and 500 Select credits per month [3][4]. Standard covers work where Notion picks the model — that’s Auto. Select covers work where you pick the model, plus your own Custom Agent runs [3][4].
Why split them? Because consumption rates differ wildly. Notion’s argument is that a few heavy tasks or a chatty Custom Agent could drain a single shared pool and starve everyday usage [3][4]. That reasoning is fine. The practical consequence is that you now have two budgets to watch instead of one.
Also worth noting: you no longer need to buy credits before you start building Custom Agents, which was the old add-on model [2][3]. Admins can still purchase extra credits for shared agents, high-consumption models like Fable and Astra, and anything past the included allowance [3][4]. Per-credit pricing for those purchases isn’t specified in the material I reviewed.
What existing customers need to know before Nov 2
Existing customers move to the new system on November 2, and admins get an email with details [3][4]. If your contract hasn’t renewed yet, you still switch on that date with the same allowances, and your credits get set at renewal based on your plan [3][4].
Three things I’d do before then: export your current Custom Agent run volume, note which agents run on autopilot and how often, and decide which ones should be funded from workspace credits instead of individual Select allowances. Team agents consume Select credits by default even when other people use them [3]. That’s the line item most likely to surprise you.
AI Command Center and admin controls
Not yet shipped, and no date beyond “coming soon” [1][3][4]. When it lands it should give admins usage by member, feature, and agent, alerts as the workspace approaches its limit, spending limits per member and agent, model-level permissions, and control over who can create, pause, or archive Custom Agents [3][4]. Members get their own usage view.
Until then you’re stitching visibility together manually. If you’re running more than a handful of agents, that gap is the strongest argument for waiting a cycle before scaling up.
Practical take: what I’d adopt, what I’d watch
Adopt Auto as the default for everyday work. Not because the 90% figure is proven, but because routing cheap-first with a re-route fallback is the right shape, and the downside if it’s wrong is a model-quality complaint you can measure. Pilot it on one team for two weeks first.
Hold off on long-running 24-hour agents for anything that touches production data. The durable execution story is solid, the “Review for me” safety check helps, but unattended agents spending credits overnight on a task nobody’s watching is a cost and correctness risk at the same time. I’d cap those runs until the Command Center ships with real spending limits.
Skip BYOK unless you’re already on Enterprise and have a provider relationship worth reusing. The alpha is narrow and undocumented. The ChatGPT connection is more interesting for individuals than for teams right now, given it doesn’t reach Custom Agents [3][4].
And do the boring admin work now. Audit agent usage before November 2, decide who pays for shared agents, and set expectations that the included credits are an estimate-friendly allowance, not a guarantee — Notion’s own claim that most users won’t run out is exactly that, an estimate [3][4]. The Notion AI rebuild rewards teams that measure. It punishes the ones that assume.