Copilot Cowork is billed in Copilot Credits at US$0.01 each on pay-as-you-go, on top of a Microsoft 365 Copilot licence, and the model you pick is the biggest cost lever you control. When I ran the same task on three models, the bill ranged from US$3.98 to US$10.69, and on a heavier job in July the dearest model used 11.7 times the credits of the cheapest.
The picker has changed since those tests. Microsoft’s model list, last updated 23 October 2026, no longer includes Claude Sonnet 4.6 or Claude Opus 4.8, the two cheapest models in my June run. The tests below still show how far model choice moves the bill. For the newer models I haven’t tested, I say so rather than guess.
How is Copilot Cowork billed?
Your Copilot licence gets you in the door, and the work itself is metered. Microsoft’s Copilot Credits Guide (October 2026) lists what Cowork needs: a Microsoft 365 Copilot licence, Cowork enabled, and usage-based billing for Copilot Credits turned on.
- The rate. Pay-as-you-go is US$0.01 per Copilot Credit. Organisations that buy a Copilot Credit Pre-purchase Plan (P3) get a discount, from 5% at 300,000 credits up to 20% at 300 million.
- What gets counted. Model responses, tool and skill calls, image generation and browser tasks all count towards consumption, according to Manage Copilot Cowork for your organization. The Credits Guide groups the cost into four buckets: models, runtime, context and tools.
- Microsoft’s rough sizes. The guide estimates 100 to 300 credits for a light task, 300 to 700 for a medium task and more than 700 for a heavy one. At pay-as-you-go rates that’s US$1 to US$3, US$3 to US$7, and over US$7.
- Who can use it. Access comes from a spending policy in the Microsoft 365 admin center that selects Cowork. Admins set limits per user or group under Copilot > Cost management, and a policy with a one-credit limit still grants access, so the way to keep someone out is to leave them out of every Cowork policy.
- Effort costs too. The effort dropdown next to the model picker defaults to Medium, and Microsoft’s model guidance says higher effort is slower and uses more credits.
Of those inputs, the model is the one you pick on every task. My Copilot Cowork guide covers the licence, setup and what other jobs have cost me.
Which models are in the Copilot Cowork picker now?
The picker defaults to Auto, which chooses from the models your organisation has enabled. As of Microsoft’s Choose a model for Copilot Cowork page (updated 23 October 2026), the options are below, with Microsoft’s own label and what I’ve measured in Cowork.
| Model | Microsoft’s label | What I’ve measured |
|---|---|---|
| Auto | For most day-to-day work | Not compared |
| GPT 6.1 Sol | Intelligent and efficient for hard work | Not tested |
| GPT 6 Astra | Latest model for tough problems | Not tested in Cowork |
| GPT 5.6 Terra | Balanced effort for common tasks | 1,781 credits on the July job |
| GPT 5.5 (Frontier) | Capable model for medium effort work | 1,069 credits in June, 1,992 in July |
| Claude Opus 5.5 | For complex, high stakes work | Not tested (I tested Opus 4.8) |
| Claude Sonnet 5.5 | Efficient for everyday tasks | Not tested (I tested Sonnet 4.6 and Sonnet 5) |
| Claude Fable 5.1 (Frontier) | Most advanced model for ambitious projects | Used for my SharePoint CRM app build, US$46.54, no comparison run |
| Grok 4.7 | Strong on long documents, coding and multi-step tasks | Not tested |
What you see depends on your tenant. Admins can turn off the Anthropic model family in the Microsoft 365 admin center, and in the EU Data Boundary only the GPT models are supported. Cowork remembers the model you chose on that device until you change it.
The June test: one task, three models
I ran the identical Cowork task three times on three models and asked for the same outputs each time. The cheapest run cost US$3.98 and the dearest cost US$10.69.
The task started from a CSV of 2025 sales figures and asked for three things: a multi-tab Excel workbook, a board-ready PowerPoint deck and an interactive HTML dashboard. After each run I used the /cost command to read the credits consumed.
| Model | Credits | Cost (US$, pay-as-you-go) |
|---|---|---|
| Claude Sonnet 4.6 | 398 | $3.98 |
| Claude Opus 4.8 | 477 | $4.77 |
| GPT 5.5 | 1,069 | $10.69 |
GPT 5.5 cost 2.7 times as much as Sonnet for equivalent work. Opus sat surprisingly close to Sonnet, about 80 cents more for the deeper reasoning.
The output quality was equivalent across all three runs. I used a custom skill called interactive-dashboard that holds the output design constant, so the difference in credits comes from the model rather than from different work being done. Of these three, only GPT 5.5 is still in the picker.
What the picker labels don’t tell you
In June the picker described GPT 5.5 as “versatile across task types”, which reads like the safe middle option. It cost about the same as running Sonnet and Opus combined.
The labels have been rewritten since, and none of them mentions cost. The October Credits Guide names the model as one of the four cost buckets but publishes no per-model rate, so the label ends up doing most of the steering.
I treat the picker as a behaviour design decision. If your people pick whichever label sounds safest, you can spend two to three times what you need to on work a cheaper model handles just as well. Auto is a sensible default if you don’t want people choosing at all. If you want predictable everyday cost, measure two or three models on your most common job and set the cheapest one that does it well as the team’s default.
This is the same pricing trap that catches teams with licensing, one layer deeper. I covered the licensing side in Microsoft 365 Copilot vs Copilot Chat.
Update: what the July tests showed
The model list kept moving after the June test, so I kept testing. In July I ran two more comparisons on a heavier job: ticketing data in a SharePoint site, with an interactive HTML dashboard, a PowerPoint deck, a Word summary and a drafted email from one prompt.
Claude Sonnet 5 against Claude Fable 5 came in at 912.2 credits and 1,738.4 credits. Fable 5 added a sticky header on the dashboard and a bit more colour, which were real differences, but I couldn’t see 826 credits’ worth of them.
Then GPT-5.6 arrived in Cowork, so I ran the same prompt across all six models available at the time:
| Model | Credits |
|---|---|
| Claude Opus 4.8 | 798 |
| Claude Sonnet 5 | 1,175 |
| GPT-5.6 Terra | 1,781 |
| GPT-5.5 | 1,992 |
| Claude Fable 5 | 2,128.5 |
| GPT-5.6 Sol | 9,324.7 |
The most expensive run cost 11.7 times the cheapest for comparable output. Opus 4.8 produced the best dashboard and was also the cheapest on this job, which is the opposite of what the June test would lead you to expect. Terra beat GPT-5.5 on both cost and quality.
Two of those six are still in the picker as of October: GPT 5.6 Terra and GPT 5.5. Sol is now listed as GPT 6.1 Sol, and Opus and Sonnet have moved to 5.5, so read the July numbers as a measure of how far apart models can land on the same job.
Which Copilot Cowork model should you use?
Start on Auto. Microsoft recommends it for most work, and it picks from the models your organisation has enabled. It works well paired with a custom skill that fixes the shape of the output.
Pick a fixed model only after you’ve measured it. Between June and July the cheapest model on my tests changed: Opus 4.8 was second in June and cheapest in July. Run your most common job on Auto and two named models, check /cost after each, and compare the output side by side.
Keep GPT 5.5 off the default. It’s the one model I’ve tested twice that’s still in the picker, and it was the dearest in June and fourth of six in July. On the July job, Terra did better work for fewer credits.
Test the new models before you trust their labels. I haven’t cost-tested Claude Opus 5.5, Claude Sonnet 5.5, GPT 6.1 Sol, GPT 6 Astra in Cowork or Grok 4.7. The labels suggest Sonnet 5.5 for everyday work and Opus 5.5 for high-stakes work, and on my tests the label didn’t predict the bill. In June the model labelled for everyday tasks was cheapest. In July the one labelled for complex, high-stakes work was.
How to keep Copilot Cowork costs under control
Copilot Credits are a new kind of budget line, and the model is the lever that decides what the meter reads. Run the same prompt on two models, check /cost after each, and you’ll quickly build a feel for what your typical tasks cost. Admins can then set per-user or per-group limits in Copilot > Cost management based on what your own jobs actually used.
The prompt moves the meter too. I built a board deck from a CSV of 2025 sales figures three ways: a loose prompt produced an eight-slide deck for 327 credits, a prompt naming four points and asking for a five-slide deck with one slide each used 165.5 credits, and the loose prompt with Work IQ searching the site for the file cost 413.7 credits. Name the outcome, the sources and the output constraints before you run the task.
Some recurring work doesn’t need Cowork at all. A policy review I ran in Cowork used 248 credits, and I moved that job into Copilot in SharePoint, where everyday work is included in the licence.
Frequently asked questions
How much does Copilot Cowork cost? Copilot Cowork needs a Microsoft 365 Copilot licence, and each task then spends Copilot Credits at US$0.01 each on pay-as-you-go. Microsoft’s Credits Guide estimates 100 to 300 credits for a light task, 300 to 700 for a medium one and more than 700 for a heavy one. My sales workbook, deck and dashboard task cost between 398 and 1,069 credits depending on the model.
How is Copilot Cowork billed? Through usage-based billing in the Microsoft 365 admin center. An admin creates a spending policy that selects Cowork, chooses pay-as-you-go through an Azure subscription, a Copilot Credit Pre-purchase Plan (P3) or existing prepaid capacity, and sets limits per user or group. P3 discounts run from 5% at 300,000 credits to 20% at 300 million credits.
What models are available in Copilot Cowork? As of 23 October 2026, Microsoft lists Auto, GPT 6.1 Sol, GPT 6 Astra, GPT 5.6 Terra, GPT 5.5 (Frontier), Claude Opus 5.5, Claude Sonnet 5.5, Claude Fable 5.1 (Frontier) and Grok 4.7. Your list depends on what your organisation allows: admins can turn off the Anthropic models, and tenants in the EU Data Boundary only get the GPT models.
Which Copilot Cowork model should I use by default? Auto if you’d rather not think about it. Otherwise, test your most common tasks on two or three models and pick on evidence, because the cheapest model changes with the task. In June, Sonnet 4.6 was cheapest on my test. In July, on a heavier job, Opus 4.8 was cheapest and produced the best output.
How do I check what a Cowork task costs?
Type /cost in the task window after a run finishes. Cowork shows the number of Copilot Credits the conversation has used so far. Admins can see consumption by user, group and service on the Consumption tab under Copilot > Cost management.