Which agent will you actually use?
ChatGPT, Gemini, Claude, Meta and now Notion all want to be your agent. Start with how people use them today.
Pew Research Center's Americans and AI 2026 (published 17 June 2026, survey of 5,119 US adults in February) found that 49% of US adults now use AI chatbots, up from 33% in 2024, and about a quarter use them daily. ChatGPT reaches 44% of adults, Gemini 24%, Copilot 17%, Meta AI 14% and Claude 6%.
Menlo Ventures' 2026 State of Consumer AI (published 16 September 2026, 5,067 US adults surveyed with Morning Consult in July) adds the numbers that matter for agents:
64% of US adults use AI, and 25% use it every day.
The average user now uses 3.0 general AI assistants, up from 2.2 a year ago.
People have given agents access to their email (36%), web browsers (33%) and messaging apps (31%).
Put those together and a pattern shows. People chat in three assistants, but fewer than a quarter run an agent regularly, and an agent is only useful once it has the keys to your apps. Every extra agent means reconnecting your accounts, re-teaching your context and paying for another plan.
So here is the rule we use: chat wherever you like, but give your agent work to one tool, and pick the one that lives where your work already lives. If your clients, content calendar and SOPs sit in Notion, that's Notion. The rest of this review tests it against that rule.
What it is
On 19 October 2026 Notion released rebuilt Notion Agents and Auto, a model router. Auto "reads each task and sends it to the model that best balances quality, speed, and cost". It is not Notion's own model, and it is on by default.
Who it's for
Solo builders and small teams who already treat Notion as their system of record: client databases, content pipelines, meeting notes, SOPs. The full agent comes with the Business plan.
What it actually does
Long-running tasks. Notion says agents can now work on a single task for more than 24 hours, up from five minutes. That is long enough to hand it an afternoon of admin.
Its own computer and browser. Agents get a computer and browser, and Notion says they can write short scripts for steps that run better as code.
Cheaper routing, by Notion's numbers. Notion claims that across 20+ common use cases, Auto delivered similar results to Opus 5.5 and Astra at about 90% lower average cost, so credits "go about 10× further". These are Notion's own tests with no published method. You also can't see which models Auto used on a task.
Approval modes. Three settings: Always ask, Review for me (the default, where a separate model checks whether an action is safe) and Skip all approvals.
More connections. Notion says agents now reach hundreds of tools through MCP and connectors, up from a few dozen. You can also connect a ChatGPT Plus or Pro plan to your personal Notion Agent.
A door for other agents. The overhauled Notion MCP lets outside agents use Notion as shared memory, and everyday read and write actions through Notion MCP and the API don't use credits. This matters for the one-agent rule (more below).
Pricing (as of 25 October 2026)
Plan | Price per member/month | AI included |
|---|---|---|
Free | $0 | Limited trial of Notion AI |
Plus | $10 | Limited trial of Notion AI |
Business | $20 | Notion Agent, AI Meeting Notes, Enterprise Search (Beta), plus 1,500 Standard and 500 Select credits a month |
Enterprise | Custom | Business features plus zero data retention with LLM providers |
Prices are from Notion's pricing page, which offers "up to 20%" off for yearly billing without listing yearly prices. Custom Agents are "Free to try, then $10 per 1,000 monthly Notion credits".
How the credits work. Standard credits pay for work where Notion picks the model (Auto). Select credits pay for work where you pick the model, and for your Custom Agents. They reset monthly, nothing rolls over, and when a balance runs out that work pauses unless your admin lets you draw on paid Notion credits.
A worked example. Notion's Auto help page estimates 5, 10 and 25 credits for a simple, medium and complex task on Auto, against 50, 85 and 150 on Opus 5.5. On those estimates, 1,500 Standard credits cover roughly 150 medium tasks a month on Auto. The same 500 Select credits cover only about five or six medium tasks if you pick Opus 5.5 yourself. Beyond the allowance, one complex Opus task (150 credits) costs about $1.50 at $10 per 1,000. From 1 December, AI Meeting Notes also draws on Standard credits at about 10 per one-hour meeting. (Our arithmetic on Notion's estimates.)
The 2 November switch. New Business and Enterprise workspaces are on this system now. Existing workspaces move over on 2 November 2026, and Custom Agents created before 2 November keep drawing from shared workspace credits.
Getting started: a one-week test
Set approvals to Always ask. Change it at the top of the agent chat. Move to Review for me only once you trust what it does.
Hand it one recurring job you do now. For example: "Every Monday, review the Content Pipeline database, list anything past its due date, draft a two-line status update for each client in the Clients database, and put the drafts on a new page for me to check."
Leave it on Auto. Choose a model yourself only when the output is weak, and note when you did. Each manual pick spends scarce Select credits.
Record your credit balance on day one and day seven. Multiply by four for a rough monthly figure.
If you build a Custom Agent, set a spending cap. Owners can cap each agent, and shared agents bill the owner's Select credits by default.
Alternatives: when to pick each
ChatGPT Work. Pick it if ChatGPT is already your daily assistant and your work runs through Gmail and Slack. OpenAI launched Work on 9 July 2026 as an agent that can "take action across your apps and files". Its help page says it can run on a schedule or on triggers such as new Gmail or Slack messages, though not on Free or Go. ChatGPT's pricing page didn't show plan prices when we checked, so we've left them out.
Claude. Pick it if your work lives in files and folders on your computer. Claude Cowork works directly in folders you choose, has its own built-in browser, and runs scheduled tasks unattended. On 16 September Anthropic merged Cowork and chat into one Claude, rolling out to Pro and Max first. Claude Pro is $20 a month billed monthly. Our earlier Claude Cowork review covers it in depth.
Gemini. Pick it if you live in Gmail and Google Docs. Google AI Pro is $19.99 a month and includes Gemini in Gmail and Docs plus Chrome auto browse. Agent Mode is listed as priority access for Ultra, which starts at $99.99.
You don't have to choose all-or-nothing. Because Notion MCP reads and writes don't use credits, you can keep Notion as your workspace and point Claude or ChatGPT at it. Our Pancake review shows the same one-MCP-connection pattern for outbound sales. If you want the routing idea in another tool, Onepin routes every voiceover line to the best voice model.
Limitations
Vendor benchmarks only. The 90% and 10× figures are Notion's own tests, with no method published.
No visibility into routing. You can't see which models Auto used, and it can switch models mid-task.
Plus only gets a trial. Notion publishes no credit allowance for Plus, so you can't cost it ahead of time.
Shared Custom Agents bill the owner. A popular team agent can drain one person's Select credits.
Admin controls aren't here yet. Notion says the AI Command Center, with spending limits and usage alerts, is coming "soon".
The ChatGPT link is narrow. It works in your personal agent only, not Custom Agents. The launch post also lists Go in one place and leaves it out of the FAQ.
Verdict
Worth it if you already run your business in Notion and pay for Business. Long-running agents, a sensible default approval mode and 1,500 routed credits a seat make it the obvious one agent for a Notion-first operator. Plus users should wait until the 2 November credit switch shows what real usage costs before upgrading.