Notion AI Agents Now Run 24 Hours — Metered by Credits

PRODUCTIVITY AI

Notion has rebuilt its agent harness for the sixth time. The new version lets a single AI task run more than 24 hours — up from five minutes — pairs it with an Auto model router that Notion says is about 90% cheaper than its premium models, and, in the same breath, puts every Business and Enterprise customer on a monthly credit meter. More power, much more running room, and a meter on all of it.

Published October 20, 2026 · By Daniel Rao

File photo: the night shift, human edition — Notion's rebuilt harness now lets AI agents work a full day on one task alone. Photo: Shixart1985, via Wikimedia Commons (CC BY 2.0).

Six rebuilds. That is the number Notion itself put on its AI agent project on Monday — "We rebuilt Notion Agent harness, again. This is the sixth generation—no rest for teams building AI!" — and you can hear a little exhaustion in the word "again". It arrived attached to the biggest promise any workspace AI product has made this year: agents that work a full day on one task, from five minutes to more than 24 hours. And then, in the very same announcement, Notion did the thing that made the whole idea make sense: it metered it. Every Business and Enterprise seat now gets 1,500 Standard credits and 500 Select credits a month — tracked, topped up, and eventually governed by a coming AI Command Center. The day-long agent and the credit meter are not two stories. They are one.

The 24-hour figure is the one that will travel furthest, so let's pin it down precisely. Notion's own blog says agents "can now work on a single task for more than 24 hours, up from five minutes" — that is "more than", not "up to", and the company's blog is the canonical wording. A 288-fold jump in task duration is either the end of a real constraint or a marketing department confusing uptime for capability. For this story, we checked Notion's official blog post and the company's LinkedIn announcement — they match — and read the full secondary reports from runtimewire, technobezz and tipranks rather than rewriting a press release from a skim. The honest verdict, for now: the engineering is real, and the independence of the claims is not.

From five minutes to a full day — the 24-hour agent

Until Monday, Notion Agent cut you off after five minutes. Think about how absurd that limit was for anything ambitious: an agent that researches a competitive landscape, drafts the deck, writes the code, runs the code, and checks its own work would time out before the first coffee refill. The new harness is designed around a simple admission — long work fails most often not because the model is wrong, but because the infrastructure gives up.

So the plumbing changed. Every session now runs on a dedicated service with what Notion calls a queue-backed worker loop, so work keeps going through restarts and hiccups — that, per the blog, is "how a task can now run for 24 hours". The agents also got smarter about memory: they compact long conversations using each model provider's native compaction where possible, and large tool results get offloaded instead of flooding the context window. Anyone curious about the underlying mechanics can start with how AI agents actually work under the hood — the context-window problem this solves is the central one.

And then Notion gave its agents something genuinely new: their own computer. With computer and browser use, agents can write short scripts that are, in the blog's words, "faster, cheaper and higher quality to run as code versus with language model tokens". That is a revealing sentence. Notion is not just letting agents run longer — it is pushing them toward doing the kind of work that doesn't need a language model at all. The cheapest token, it turns out, is the one you never generate.

There is a side to long-running agents the announcement treated lightly, and it deserves more than a passing line. An agent with its own computer, a browser, and a 24-hour shift is a dramatically larger surface for things to go wrong than a chatbot that answers one question at a time. Notion's answer is a new "Review for me" mode, which uses a separate model to check whether an action is safe before the agent takes it — a watcher for the worker. Sensible, but it is a company-designed control, not an independent safety guarantee. We have seen what happens when agents go rogue in the wild; the longer the leash, the more the safety story matters. Notion gets credit for shipping a control rather than a shrug — but the control has not been independently evaluated, and neither has the agent.

Auto: the router that swaps the engine mid-flight

The second big piece is a new model router called Auto. It reads each task and sends it to the model that best balances quality, speed and cost — and if a model is not making meaningful progress, "Auto re-routes to a different model" mid-task. This is the most interesting technical idea in the whole release, because it treats model selection as a runtime decision instead of a settings-page decision. Most products ask you to pick your model; Notion is betting you shouldn't have to think about it at all.

Then come the numbers, and here we must be careful, because every one of them is Notion's own. The company says that across 20+ common use cases, Auto "delivered similar results to Opus 5.5 and Astra at about 90% lower average cost. That means credits go about 10× further." "Astra" and "Fable" are model names from Notion's own terminology — the announcement does not say whose models they are, and nobody else has verified the benchmark. The methodology is unpublished. We are, in other words, asked to take Notion's word that its new router is ten times more efficient than its premium options.

Shir, Notion's Head of AI and Core Product, put the logic plainly in the blog: "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." That sentence is the honest core of the announcement, and it explains the pairing of the 24-hour agent with the router. You cannot hand out day-long compute and charge like it's 2024. The router is what makes the duration affordable — if its numbers hold.

1,500 + 500: the two-tier meter

Here is where the announcement gets consequential for every buyer. Business and Enterprise seats now come with two monthly credit pools: 1,500 Standard credits for work where Notion picks the model, and 500 Select credits for work where you pick the model or run Custom Agents. Admins can buy additional credits and control how they're allocated. And Custom Agents — which used to need a paid credit purchase to even start — are now free to try: "Try Custom Agents without paying first using your included monthly Select credits."

A Notion executive's companion announcement, reposted on the company's LinkedIn page, made the industry context explicit: "The credit model is similar to what several other companies (Anthropic, OpenAI, Cursor) have adopted over the past year." Notion is not inventing metered AI here. It is adopting the industry's new standard pricing grammar — credits instead of seats — and applying it to agents that never sleep.

The rollout is already in motion: the new system is live today for new customers, and existing customers move over on November 2, with admins getting an email with the details. Notion's own estimate is that "the vast majority of users are unlikely to run out of credits" — note the phrasing, which is a prediction, not a fact, and one the company has an interest in being right about.

Here is my read, and I'll take a stance on it: this is Notion's most important bet since Custom Agents themselves, and it isn't the router or the 24 hours — it's the meter. Notion has decided that AI agents are infrastructure, priced like cloud compute, not features priced like software. Anthropic, OpenAI and Cursor already reached the same conclusion. Whoever makes agentic AI cheap, metered and governable wins the enterprise, and Notion just made the boldest public claim that efficiency — not raw intelligence — is now the product. That claim is unverified. But the direction is unmistakable, and Microsoft's own push to govern its AI agents shows the largest players are converging on the same idea from different directions.

What "more for less" actually shipped

Beyond the headline pieces, the release is a long list of real improvements. Agents run up to 50% faster on the web and mobile apps — the company's LinkedIn copy phrases it as chats loading up to 50% faster on mobile — and in Notion's own testing, Custom Agent runs cost about 30% less. (Company testing again; same caveat.) The company also says seven new models were added in the past month from Anthropic, OpenAI and open-weight providers, widening the pool Auto can route into.

Two integrations deserve attention. First, you can now use your ChatGPT Plus, Pro or Go account inside Notion — in your personal Notion Agent to start, not Custom Agents, with more plans and surfaces to follow — built, per the blog, "in close partnership with OpenAI". That is a remarkable sentence to see from a company competing with OpenAI's own products: the workspace wars are giving way to a cross-pollination era. Second, Bring Your Own Key is in alpha for Enterprise customers, which lets the most cautious IT departments keep their model spend on their own contracts.

Then there is MCP. Notion calls it "our fastest-growing product ever" on LinkedIn, and the release lets you sync memories to Notion, connecting ChatGPT, Claude, Grokbot and Muse — with Skills via MCP coming soon. One more thing is coming but not here: the AI Command Center, with credit usage by member, feature and agent, alerts near limits, spending limits, and model-selection governance. "Soon", says the blog. It is announced, not live — and we will not write about it as if it shipped.

The fine print Notion did not underline

Every announcement deserves an accounting of what it did not say, and this one has a few lines worth reading twice. Ryan Merket, reporting for runtimewire, noted that the release "provides no customer-level data on task quality or typical credit use". That is the single most important sentence in the secondary coverage: we know how many credits you get, we do not know how far they go. The 10× efficiency figure, the 30% cheaper runs, the >99.9% infrastructure uptime — all Notion's own, all unverified by anyone else. Notion deserves credit for the candor of its own testing framing; the numbers are at least presented as the company's. But a benchmark no outsider can reproduce is a claim, not a result.

There is also a naming trap in the release worth flagging plainly: "Astra" and "Fable" appear as high-credit-consumption models in Notion's terminology, and the announcement never says whose models they are. Do not go looking for them as public Anthropic or OpenAI products — in this release, they are Notion's words, and nothing more. And the AI Command Center's "coming soon" status matters more than it looks: the governance controls that make the credit system enterprise-safe — spending limits, model-selection governance, member-level views — are the announced future, not the shipped present. Admins buying credits today are buying into the meter before the dashboard exists.

The workspace agent war now has a price tag

Notion is not rebuilding in a vacuum. Asana launched Agentic Work Management in mid-September, pairing its AI Teammates with an AI Studio builder, 30 pre-built agents and its own credit tiers — and claimed in early October to have cut its browser agent's cost 76× with a model optimisation. Monday.com rebuilt itself around Claude, with Anthropic's own case study citing 5 million-plus agent interactions since May. Coda rebranded to Superhuman Docs and swapped Coda AI for beta-stage Docs AI. And OpenAI's Dots put personal agents directly onto Jira, Asana and Linear boards. Everybody is building the same thing: an agent layer over work.

Against that lineup, Notion's play reads as an efficiency-and-openness pitch against platform-specific ecosystems: long-running agents with their own computer and browser, model-agnostic routing, and integrations — including OpenAI's own ChatGPT — that a closed-platform vendor would never ship. The co-founder Akshay Kothari demoed the update on X on Monday, per runtimewire's account of the video post, and a company executive summed it up in the announcement reposted on LinkedIn: "You can now do much more complex things (computer + browser use!), let the AI run much longer (up to 24 hours!), and do all of this in a much more secure setup." Exclamation marks included — this is a launch, after all.

The real story of this release is not that an agent can now run for 24 hours. Plenty of agents will waste a day beautifully. The story is that Notion is betting the entire economics of workspace AI on a simple proposition: make the agents cheap enough, meter them honestly, and the long runs become normal. If Auto's numbers are real, that is a genuine breakthrough — 10× efficiency changes what enterprises let agents do unsupervised. If the numbers are optimistic, the meter is still the industry's direction of travel, and Notion got there with the loudest claim. Either way, the five-minute leash is gone. The question now is what the metered agents will do with their day — and what it will cost when they do. The answers to both, notably, are numbers only Notion has seen. We will be watching the first independent tests the way hawks watch a field mouse: that is when this story becomes real.

Sources: Notion official blog, "Agents, rebuilt to do more for less," October 19, 2026; Notion LinkedIn company announcement, October 19, 2026; runtimewire ("Notion lets agents run past 24 hours as it meters AI use," Ryan Merket, October 19, 2026); technobezz ("Notion AI Adds Two Credit Pools and Longer Agent Tasks"); tipranks ("Notion Expands AI Platform With Upgraded Agents, Routing and Unified Credits"); aiagentstore.ai daily AI-agent digest, October 20, 2026.

About the author: Daniel Rao covers AI, enterprise software and the business of automation for TechPulse USA, with a focus on agentic AI and the platforms competing to run it.