Verdict: Microsoft has launched Copilot Cowork as a generally available, agentic AI assistant for Microsoft 365, but the pricing model has changed from a flat per-seat fee to metered Copilot Credits. This makes agent usage more transparent, but also more unpredictable — especially for teams that run many long, multi-step workflows. To manage that cost, Microsoft is reportedly evaluating a Microsoft-hosted, fine-tuned version of DeepSeek V4 as a lower-cost model option alongside its existing Anthropic and OpenAI models. For most small and mid-sized businesses, the practical takeaway is simple: don't enable Cowork without first setting spending caps, modeling likely usage, and testing it on a small group of power users.
Last verified: 2026-06-17 · What changed: Copilot Cowork reached general availability on June 16, 2026, with usage-based billing in Copilot Credits. · Explore next: How Much Does AI Cost for a Small Business in 2026?
What is Copilot Cowork?
Copilot Cowork is Microsoft's agentic AI layer inside Microsoft 365. Unlike the standard Copilot chat experience, which answers a prompt and stops, Cowork is designed to carry out long-running, multi-step tasks across Outlook, Teams, Word, Excel, PowerPoint, and SharePoint.
Examples of what it can do, according to Microsoft's documentation:
- Draft, send, and manage emails and calendar events.
- Create Word documents, Excel spreadsheets, and PowerPoint decks from scratch.
- Post updates in Teams channels and search across your organization's data.
- Run background tasks even when the user's device is off.
Microsoft announced that Cowork reached general availability worldwide on June 16, 2026, after a three-month Frontier preview. The company says more than half of the Fortune 500 used Cowork during the preview, including Accenture, Capital Group, Koch, and Zurich Insurance (Microsoft 365 Blog, June 16, 2026).
Why Microsoft is moving from per-seat to usage-based pricing
Copilot Cowork's pricing reflects a broader shift in the AI industry: agentic workflows are expensive to run. A single Cowork task can pull in large amounts of company context, call multiple tools, iterate on a plan, and produce several outputs. Each of those steps consumes model tokens and compute.
Charles Lamanna, Microsoft's Executive Vice President for Copilot, Agents, and Platform, told Axios that flat-rate subscriptions become unsustainable when power users trigger hundreds of autonomous tasks every week (Axios, June 16, 2026).
Microsoft is therefore treating AI more like cloud infrastructure: you pay for what you use. The company has made a similar move with GitHub Copilot, which switched to usage-based billing in June 2026.
How Copilot Credits work
Microsoft's new billing system is built around Copilot Credits. According to the official announcement:
- Cowork requires a Microsoft 365 Copilot User Subscription License (USL) for each user — the same $30/user/month seat that powers Copilot Chat and in-app assistance.
- Cowork usage is then billed on top of that seat in Copilot Credits.
- Each task's credit cost is determined by four factors: model use, context retrieval, tool calls, and runtime.
- Microsoft classifies tasks as light, medium, or heavy based on how much work they do.
- Pay-as-you-go pricing is $0.01 per Copilot Credit. Microsoft also offers a prepaid P3 option for volume discounts.
Cowork is off by default. Admins must enable it, choose who gets access, and can set spending limits at the tenant, group, and user levels (Microsoft 365 Blog, June 16, 2026).
| Cost layer | What you pay for | Pricing |
|---|---|---|
| Microsoft 365 Copilot USL | Seat license for Copilot Chat and in-app experiences | $30/user/month (annual) |
| Copilot Cowork usage | Agentic tasks run in credits | $0.01 per Copilot Credit (PayGo) or prepaid P3 |
| Extra plugins/connectors | Some third-party plugins may carry separate fees | Varies by vendor |
Sources: Microsoft 365 Blog, Microsoft Learn — Copilot Cowork overview, Microsoft Learn — Usage-based billing overview
Where DeepSeek fits in
To keep Cowork affordable at scale, Microsoft is reportedly exploring a Microsoft-hosted and fine-tuned version of DeepSeek V4 as a lower-cost model option. Axios reported this on June 16, 2026, citing Microsoft sources (Axios).
DeepSeek, the Hangzhou-based AI lab, has built a reputation for high performance at lower inference cost. DeepSeek V4, released on April 24, 2026, comes in two sizes:
- DeepSeek-V4-Pro: 1.6 trillion total parameters, 49 billion active per token.
- DeepSeek-V4-Flash: 284 billion total parameters, 13 billion active per token.
Both models support a 1-million-token context window. At the time of release, DeepSeek priced V4-Flash at roughly $0.14 per million input tokens and V4-Pro at roughly $0.435 per million input tokens — far below leading Western frontier models (DeepSeek API docs, DeepSeek Hugging Face collection).
Microsoft is not new to DeepSeek. The company added DeepSeek R1 to Azure AI Foundry and GitHub in January 2025, positioning it as a cost-efficient reasoning model within an enterprise-governed environment (Microsoft Azure Blog, January 29, 2025).
Why this matters for enterprise trust
DeepSeek's Chinese origin has raised geopolitical and data-sovereignty questions. Microsoft's response, according to Axios and the Azure blog, is to host the model inside Azure, apply its own governance and content-safety layers, and keep customer data within Microsoft's cloud boundary. This mirrors how Microsoft already offers other third-party models through Azure AI Foundry.
If Microsoft proceeds, DeepSeek V4 would likely be one option among several in a model picker, not a replacement for Anthropic or OpenAI models. Cowork already lets users choose models in Frontier, and a future "Cowork 1" Microsoft-tuned model is also expected to lower costs (Microsoft 365 Blog, June 16, 2026).
What this means for you
If you run a small or mid-sized business already paying for Microsoft 365 Copilot, here's how to think about Cowork:
- Don't treat it as free. The Copilot seat is still required, and Cowork tasks add metered costs on top. A few heavy users can quickly inflate the bill.
- Start with spending caps. Cowork is off by default. Leave it that way until you've set tenant-, group-, and user-level budgets in the Microsoft 365 admin center.
- Model your usage before enabling broadly. Microsoft provides a downloadable cost estimator at aka.ms/CustomerCoworkEstimator. Use it.
- Test with a small group first. Pick 3–5 power users, run typical workflows, measure credit consumption, and compare the output quality against the cost.
- Watch for the cheaper model tier. If Microsoft adds DeepSeek V4 or Cowork 1 as a lower-cost option, it could make routine, repetitive agent tasks much more affordable — but only if the quality meets your needs.
- Compare with alternatives. If your main use case is customer service, HR, or IT automation outside Microsoft 365, dedicated agent platforms may be more cost-predictable. See Best AI Customer Service Tools for Small Business 2026 and How Much Does AI Cost for a Small Business in 2026?.
Related reading
FAQ
Do I have to pay extra for Copilot Cowork?
Yes, beyond the standard Microsoft 365 Copilot seat. Cowork usage is billed in Copilot Credits on a pay-as-you-go basis at $0.01 per credit, or through a prepaid P3 commitment.
How much does a typical Cowork task cost?
Microsoft groups tasks into light, medium, and heavy patterns. Light tasks use the fewest credits; heavy tasks — broad aggregation, deep reasoning, many outputs — use the most. The exact credit count depends on the model, context, tools, and runtime.
Is DeepSeek already inside Copilot Cowork?
Not yet. As of June 2026, Microsoft is reportedly evaluating a Microsoft-hosted, fine-tuned DeepSeek V4 option. It has not been officially announced as a live Cowork model. Treat this as reported exploration, not a shipped feature.
Is my data safe if Microsoft uses a Chinese-origin model?
Microsoft says any DeepSeek deployment would be hosted within Azure, governed by Microsoft's enterprise security and compliance layers, and subject to the same data-handling commitments as other Azure AI Foundry models. Organizations with strict data-sovereignty requirements should review their Azure contracts and compliance posture before enabling third-party models.
Can I disable Cowork to avoid surprise bills?
Yes. Cowork is off by default. Admins control enablement, access, and spending limits. You can also set hard caps and receive usage alerts.
Should a small business turn on Cowork now?
Only if you have a clear use case, a budget, and a plan to measure ROI. For most small businesses, the safest path is a limited pilot with spending caps, not a company-wide rollout.
Discussion
0 comments