Notion Custom Agents vs Copilot Studio vs Glean: Cost per Agent Run

A weekly status report agent costs $0.08 to $0.18 a run on Notion, by Notion's own estimate. On Copilot Studio the same job lands near $0.30 at list rates, and on Glean it cannot be priced in dollars from public documents at all. Notion custom agents pricing is the cheapest per run of the three. It is not automatically the cheapest contract, and each platform runs the agent as a different person.

The weekly reporting agent we priced on all three platforms

Usage pricing is hard to compare because each vendor meters a different unit. Notion meters credits drawn by a whole run. Microsoft meters each action inside a run. Glean meters a run but publishes a percentile range instead of a rate. The only fair comparison is to hold one job constant and price it on each meter.

Every Monday at 8am, an agent reads a project tracker, reads the past week's task updates, writes a one-page status summary, and posts it to the team channel. It runs 52 times a year. Nobody is watching when it runs.

This post is written for the head of operations or delivery who owns that report and will have to defend both its bill and its access to an admin.

Two things are outside scope. I do not compare answer quality, because a status summary is not a hard reasoning task. I do not compare search and indexing, which is covered in the analysis of how Glean and Copilot sync permission changes.

Notion Custom Agents pricing after the 2 November switch

Notion Custom Agents were already funded by purchased Notion credits. What changes on 2 November 2026 is who pays first. Notion's help centre states that new Business and Enterprise workspaces moved to the new credit system from 19 October, and existing workspaces move on 2 November 2026.

The rate card Notion publishes

Notion's pricing page lists Custom Agents as free to try, then $10 per 1,000 monthly Notion credits, available only on Business and Enterprise plans. Unlike either competitor, its help centre prices example agents per run, in dollars, from current usage data.

A status update Custom Agent is estimated at $0.08 to $0.18 per run, or 57 to 133 runs per 1,000 credits. A daily brief agent runs $0.10 to $0.30. A Q&A agent runs $0.03 to $0.11. Notion's own worked example is a status agent run 60 times a month for about $4.80 to $10.80.

Applied to our weekly report, 52 runs cost $4.16 to $9.36 a year. That is the whole bill if the workspace buys credits for it. The spread between the low and high estimate is 2.25 times, the narrowest of the three platforms.

Notion names four drivers of credit use: how much the agent reads, how many steps it takes, how often it runs, and which model it uses. A report that scans a large database every week will sit at the top of the range, so trimming the tracker view the agent reads is the cheapest optimisation available.

Select credits change who pays for the run

Under the new system, each member gets 1,500 Standard credits and 500 Select credits a month, with no rollover. Select credits cover work where the member picks the model or runs their own Custom Agents.

A Custom Agent now draws, in order, from the owner's Select credits, then any extra allowance an admin has given that owner, then the workspace's shared Notion credits if an admin allows it. Agents created before 2 November keep drawing from shared workspace credits.

Our report uses 8 to 18 credits a run, so 32 to 90 credits in a four or five run month. That is under a fifth of one member's 500 Select credits. For a new agent, the marginal cash cost is plausibly zero, because the owner's allowance absorbs it. I think this is the most consequential pricing change of the three.

There is a catch in the drawdown order. The owner's Select credits are shared with everything else that owner does with a chosen model. A heavy user who also owns five scheduled agents will exhaust the allowance mid-month, and the agents then fall through to workspace credits or pause.

Copilot Studio pricing for an autonomous weekly agent

Copilot Studio prices in Copilot Credits. Microsoft's licensing guidance lists $200 per pack of 25,000 credits a month, billed annually, or $0.01 per credit pay-as-you-go. The pack works out to $0.008 a credit, about 20% below the meter, on credits that expire at month end.

Counting the credits in one run

The billing table on Microsoft Learn charges 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action, 10 for tenant graph grounding, and 13 per 100 agent flow actions. Reasoning models add 10 credits per 1,000 tokens on top of the feature rate.

Microsoft's own worked example is an autonomous agent triggered by each new order that takes four actions, billed at 4 times 5, or 20 credits. Our report has the same shape: read the tracker, read the task list, post the summary, plus one more action to format it. Add one grounded search over SharePoint for context and the run costs 30 credits.

At $0.01 a credit that is $0.30 a run and $15.60 a year. On spare pack capacity it is $0.24 a run. Switch the summary to a reasoning model that consumes 8,000 billed tokens and the run gains 80 credits, reaching $1.10.

You can price a run before it happens by reading your agent's action list. But every extra tool call you add for reliability shows up on the bill at 5 credits each, a pattern the comparison of per-task meters across agent vendors traces across vendors.

Why the Microsoft 365 Copilot licence does not cover it

Most buyers assume their Microsoft 365 Copilot seats make agents free. For interactive use, they largely do. Microsoft's billing table marks every line above as no charge when used by a Microsoft 365 Copilot licensed user.

The footnote carries the condition. The inclusion applies when the agent "operates using the authenticated Microsoft Copilot USL user's identity". A scheduled agent does not. Microsoft's triggers documentation states that event triggers can use only the agent author's credentials. The same billing page says agent flows called by an autonomous agent consume credits even if your users hold Copilot licences.

My reading is that a Monday 8am report is billable at the rates above regardless of how many Copilot seats you own. If your Microsoft account team reads it differently, get that in writing, because it is the single line that decides whether this agent costs $0 or $15.60 a year. The wider map of what Microsoft now meters is in the breakdown of Copilot credits and metered billing.

Glean agents pricing: a credit range without a dollar rate

Glean sells Enterprise Flex as a per-user licence plus an organisation-wide pool of FlexCredits. Glean states that the pool is included with the licences and that you pay only above the initial allocation. Agent creation, testing and sharing are unlimited. Agent runs consume FlexCredits at the current rate card.

The 50th and 90th percentile run

Glean's rate card does not give a fixed rate for an agent run. It marks the line Variable and publishes typical consumption: about 7 FlexCredits at the 50th percentile and about 114 at the 90th. For comparison, a premium Thinking Mode query runs about 35 and 120.

A weekly report that behaves like a median run uses 364 FlexCredits a year. One that behaves like a 90th percentile run uses 5,928. The ratio between them is 16.3 times, against 2.25 times for Notion's range. A finance team can budget around a 2 times spread. A 16 times spread needs a usage cap, not a forecast.

The dollar figure is the gap. None of Glean's pricing pages publishes a price per FlexCredit or for a FlexCredit pack. Glean's Core Suite plan is the exception: it bills model usage at provider API rates, listing Claude Sonnet 5.5 at $2.00 per million input tokens and $10.00 per million output tokens. A report reading 40,000 tokens and writing 2,000 would cost about $0.10 on that plan, an illustrative figure built from my token assumption.

Per-user schedules multiply the runs

The bigger cost driver on Glean is structural. Glean's schedule documentation says "scheduled agents are activated by individual users and run according to each user's schedule". Bulk scheduling runs each execution "in the user's own context with their own permissions".

So if 12 team leads each subscribe to the Monday report, Glean runs it 12 times. On Notion and Copilot Studio one scheduled run posts one summary to a shared channel. On Glean, 12 subscribers at the median rate use 4,368 FlexCredits a year, and at the 90th percentile 71,136. Before you compare per-run rates, count how many people will schedule the agent.

AI agent cost per run, side by side

The table holds the job constant and applies each vendor's published meter. The Copilot rows use my 30-credit run built from Microsoft's table. The Glean row uses Glean's percentiles and stays in credits, because that is all Glean publishes.

Where Notion wins and where it does not: it has the lowest published run price and the only one a vendor states in dollars. It does not win when the report must read Microsoft 365 data that already sits under Copilot, because connecting it through a third-party tool adds a credential problem described below.

Every figure here is small. At under $60 a year, the run price of one weekly agent does not decide the platform. The pack minimum, the number of subscribers and the identity model do, and those are the same line items examined in the guide to why AI credits and tokens confuse buyers.

Who the agent is when it runs unattended

A scheduled agent has no user sitting in front of it, so each vendor has to decide whose access it carries. For a weekly report that touches project data, this matters more than the $10 difference in annual cost.

Notion: explicit grants plus borrowed credentials

Notion's help centre states that "agents act only on the pages, databases, and external apps you explicitly grant access to" and never have full workspace access by default. That is the cleanest default of the three for Notion content.

External tools work differently. Each MCP connection is unique to one agent and "uses the credentials of the person who authenticated it". Notion also states that anyone with access to the agent can see its outputs, even without access to the underlying private Slack channel.

So a report that pulls from Jira through MCP reads Jira as whoever connected it. If that person can see a confidential project, the summary can describe it to everyone who can view the agent.

Copilot Studio: the maker's credentials, by design

Microsoft is explicit. "Event triggers can only use the agent maker's credentials for authentication", and for an agent to run autonomously, every action that needs authentication must use maker credentials. Microsoft adds that this "might allow users of an agent to use the agent to access data and systems using that same authorization".

So the identity of the person who built the Monday report is a security decision. If a senior manager builds it with their own account, the agent reads with a senior manager's reach. The fix is a dedicated maker account with narrow scopes, the pattern laid out in the comparison of agent credentials against user impersonation.

Glean: the person who scheduled it

Glean takes the opposite position. Its security guidance states that "agent access is enforced at execution time based on the executor's identity (the signed-in user running the agent), regardless of who created or published the agent". Background agents can only use tools admins have cleared to run without confirmation.

This is the safest default for data exposure, and it is the reason Glean's runs multiply. Each subscriber gets a report built from what they can see, and each one is billed. Glean's Independent Agents, in beta, add an agent-owned identity scoped by admins for shared workflows.

Spend controls: what stops a runaway weekly agent

A report someone edits to run hourly, or to read the whole workspace, can drain a shared pool quietly.

Notion notifies owners at 80% and 100% of workspace credit usage. When credits run out, agents on shared workspace credits pause automatically until credits reset or an admin adds more. Pausing is the right failure: a missed Monday report is visible, and an overage is not.

Glean lets admins set monthly limits for the organisation, per user, per agent, per department or per group. Alerts go out at 50%, 75%, 90% and 100%. Blocking is optional, and Glean says it lets the in-flight action finish before blocking further billable use, so a small overage can occur.

Copilot Studio enforces prepaid capacity at 125% of what you bought, and Microsoft notes that users with a Microsoft 365 Copilot licence might not be blocked. It also states that a monthly limit set on an agent does not stop the agent flows it calls. For a scheduled agent built on flows, the per-agent limit is weaker than it looks. Spend caps as a category are compared in the review of HubSpot's AI spend caps.

Where this comparison is weakest

The Copilot run is my construction. Microsoft publishes rates per feature, not a cost for a status report, so the 30-credit run is built from its table and its own order-processing example. A real agent built by a different maker might use six actions or two. The Notion figure, by contrast, is Notion's own estimate, and estimates published by the vendor tend to describe a well-built agent rather than a typical one.

The Glean comparison is incomplete by necessity. Without a public price per FlexCredit, I can show the shape of the cost but not its size. Glean's pool is described as generous and included with seats, so for many buyers the marginal cost of one weekly agent is genuinely zero until the pool runs dry. That could make Glean the cheapest of the three in practice, and I cannot rule it out.

The deepest weakness is that run price is the wrong unit for most buyers. The cost that decides this purchase is the seat or platform underneath. Notion credits require a Business or Enterprise plan. Copilot Studio packs come in $200 monthly blocks billed annually. Glean sells seats with a pool attached. A per-run comparison is honest about the meter and silent about the minimum.

The source base is vendor documentation, written by the party with an interest. I have relied on the pages where each vendor states an inconvenient fact about itself, such as Microsoft's warning about maker credentials and Glean's 16 times spread. No independent measurement of agent run costs on any of these platforms appears to be published.

Frequently asked questions

Where to start

Before Monday, open the admin view on whichever platform you use and find the scheduled agents that already exist. For each one, write down two things: whose credentials it reads with, and how many times a week it actually runs. On Copilot Studio, check the maker account. On Glean, count subscribers. On Notion, check who authenticated each MCP connection.

Then move any agent that reads with a senior person's credentials onto a dedicated, narrowly scoped account. That change costs nothing and removes the larger of the two risks in this post. The scope review in the guide to reviewing agent OAuth scopes is a sensible template for the audit.

Related reading