Copilot Cowork went generally available on June 16 with usage-based billing. Cowork now charges Copilot Credits on top of your Microsoft 365 Copilot seat, a penny per credit. Copilot Chat, already included in that seat, is still unlimited for your $30 per Month license.
For years everyone just reached for the most powerful option, because there was not cost to it. Now, the game is going to shift to those being efficient and effective with AI automation.
Estimation is difficult to do before actually run the task.
The natural move is to get ahead of the bill by estimating. Work out what a task will cost, then decide whether to run it. That doesn’t actually work every time.
A task's cost comes from four things: the model you pick, the context the agent retrieves, the number of tool calls it makes, and how long it runs. You set the model before the work starts so you have control there. The other three only come into existence once the agent is already running. Nobody can tell you up front how many tool calls a job will take or how much context it will pull, and that includes the model itself. The number is unknowable until the run is finished. It’s also not guaranteed to be consistent from run to run.
So the control has to come from somewhere other than a forecast. It comes from testing. Pick the cheapest Microsoft model that does the job, and scope the task so it can't run away from you.
Work down the ladder, stop at the first rung that works
This starts with a decision ladder.
Copilot Chat first. Included in your seat, no credits. Quick questions, single-document summaries, drafting an email, a lookup. A large share of what people open Cowork for never needed Cowork. Chat handles it for no added cost.
Researcher next. When the job is real multi-source research and reasoning across a body of material, Researcher is built for that. Reach for it before you reach for an agent.
Cowork last. Open it when the work is multi-step, spans several tools, runs long, and ends in a finished deliverable. That is the job Cowork is built for. Opening it by reflex for smaller work is how the bill creeps up on you.
Inside Cowork, the model picker is a cost control
The model picker is a cost control you have actual control over, and most people leave it parked on the strongest option. As of now, Opus 4.8 is the most capable and the most expensive. Sonnet 4.6 is cheaper and faster. Lighter models cost even less. It is now important to learn the relative strengths and weaknesses of each model so you can get the job done and reduce costs.
Match the model to the job. A heavy reasoning task, planning, and complex content gets Opus. Reformatting a list, pulling figures from a handful of files, routine drafting, all of that runs fine on Sonnet at a fraction of the price. Defaulting to the top model on every task is the most common way people overspend without noticing.
Focus on the Scope
Of the four cost drivers, scope is the one in your hands, and the one that runs away fastest. "Summarize this folder" and "go through everything and tell me what matters" can chew up a lot and get you to the same answer at very different prices. The open-ended version sends the agent crawling, retrieving, and calling tools well past the point of useful return.
Name the files. Name the date range. Name the output you want. A tight scope is the single biggest thing you can do to be efficient.
Two safety nets
Scoping catches most of it. But not always
Type /cost mid-task to see real spend while the agent works. If the number is climbing past what the job is worth, you find out during the run instead of after it.
Set a hard cap with auto-stop before you run anything you are unsure about. The task halts at the ceiling instead of sailing through it. On a heavy job against a large data set, set the cap first, every time. This is important for scheduled, unattended, tasks.
Run this before you commit a task to Cowork
Here is the habit that ties it together. Before you hand a task to Cowork, paste this into Copilot Chat, which costs you nothing, and let it tell you the cheapest surface for the job.
You are my Microsoft Copilot cost advisor. Recommend the cheapest
Microsoft solution that does the job well, rather than the most powerful one.
THE TASK:
[what you want done, one or two sentences]
WHAT IT TOUCHES:
[the data: one Word doc, my last 20 emails, the whole SharePoint site,
or nothing if you're just drafting]
HOW THIS WORKS:
- Copilot Chat is included in my seat and costs no credits. Best for
quick Q&A, single-doc summaries, drafting, lookups.
- Researcher is for deep, multi-source research and reasoning.
- Cowork costs credits and is for multi-step, multi-tool, long-running
work that produces a finished output.
- Cowork credits come from four things: model used, context retrieved,
number of tool calls, and runtime. Billed at $0.01/credit. Tasks are
light, medium, or heavy.
- Within Cowork, model is a cost lever: Opus 4.8 is most capable and
most expensive, Sonnet 4.6 is cheaper and faster, lighter models
cheaper still.
TELL ME:
1. Which solution, and why.
2. If Cowork: which model, and the likely tier (light/medium/heavy).
3. A rough credit range and dollar range for that choice.
4. The single biggest cost driver in this task.
5. How to reshape the task to make it cheaper, if it can be.
6. A spending cap to set before running it.
The one habit to build is a five-second pause before you open Cowork. Ask whether Copilot Chat or Researcher already does the job. If the answer is genuinely Cowork, match the model to the work, scope it tight, and set a cap before you run it.
If you want help setting credit caps, training your team on the habit, or modeling what Cowork will cost across your tenant before you turn on Cowork for users, reach out.
Topics: Copilot Cowork, Microsoft 365, Copilot, cost management, usage-based billing