Copilot Cowork pricing: the agent-cost signal

Last verified: 2026-06-16.

In short: Microsoft’s Copilot Cowork launch is a pricing signal for enterprise AI agents. Long-running agents do not behave like chatbots: they retrieve context, call tools, run for longer, and can burn through model calls while they complete work. Microsoft is therefore moving Cowork to usage-based billing, and Axios reports that Microsoft is exploring a Microsoft-hosted DeepSeek V4 or another open-source model as a cheaper option. Treat the DeepSeek part as reported exploration, not a confirmed product choice.

Microsoft announced that Copilot Cowork is now generally available worldwide on June 16. The company describes Cowork as an agentic system for complex, long-running, multi-tool tasks across Microsoft 365. In the same announcement, Microsoft says Cowork requires a Microsoft 365 Copilot User Subscription License and is then billed on a usage basis in Copilot Credits.

That matters more than the launch label. A normal chat assistant can often be priced as a per-seat bundle because the product has a relatively predictable interaction pattern. A work agent is different. It may search company context, call tools, produce several outputs, wait for checkpoints, and keep working after the first prompt. Microsoft’s pricing model is a public reminder that agent products need a real cost model, not just a seat count.

What Microsoft actually announced

The primary-source facts come from Microsoft’s own post. Microsoft says Copilot Cowork is generally available worldwide after a Frontier preview, and says more than half of the Fortune 500 used Cowork during that preview. Microsoft positions Cowork as a system that executes complex, long-running, multi-tool tasks rather than returning only a draft or recommendation.

The pricing detail is the important part for buyers and builders. Microsoft says Cowork usage is billed in Copilot Credits, and that the price for each task is calculated from four inputs: model use, context retrieval, tool calls, and runtime. Microsoft also says it observed light, medium, and heavy task patterns during Frontier usage.

That is a useful mental model for evaluating any enterprise agent. If the vendor only talks about seats, ask what happens when a user starts hundreds of tasks, each with retrieval, tool calls, and long-running execution. If the vendor only talks about model tokens, ask how context retrieval, connectors, and tool runtime are counted. Cowork’s public pricing language gives teams a checklist for the cost parts that are usually hidden.

Microsoft also says its newest model, Cowork 1, will be a secure, fine-tuned model released in the coming weeks and “post-trained to handle tasks at a substantially lower cost.” That is Microsoft’s claim, not an independent benchmark. But the direction is clear: agent platforms will route work across models based on task fit and cost, not simply send every step to the most expensive frontier model.

Where DeepSeek fits — and where it does not yet fit

Axios’ Ina Fried reports that Microsoft is considering a Microsoft-hosted version of DeepSeek as a cheaper model option for Copilot Cowork. Axios says Microsoft is exploring a fine-tuned DeepSeek V4, or another open-source model, as a lower-cost alternative to the Anthropic and OpenAI models now powering Cowork.

The wording matters. Microsoft has not publicly said, in the GA announcement, that DeepSeek is shipping inside Cowork. Axios reports that Microsoft expects to make a lower-cost model available in the coming weeks and confirm the choice then. So the safe reading is: Microsoft is publicly moving Cowork to usage-based pricing; Axios reports DeepSeek V4 or another open-source model is under consideration for cheaper routing.

Axios also reports that, if Microsoft goes forward with DeepSeek, the model would be optional and fully hosted on Azure, keeping customer data within Microsoft’s cloud and covered by Azure enterprise security, compliance, and data-residency controls. That is the enterprise wrapper. The model’s origin will still draw attention, but the product question for buyers is narrower: what is the data boundary, what model is actually invoked, who can choose it, and how are outputs logged and reviewed?

The real story is not DeepSeek. It is agent economics.

The DeepSeek name is the attention hook. The durable story is cost pressure. Agentic tools can keep calling models as they work through a task. Axios quotes Charles Lamanna, Microsoft’s executive vice president for Copilot, agents and platform, saying testing showed Cowork could not be offered on an unlimited-use basis because some users do hundreds of tasks a week and costs can go very high.

That is the same problem any serious agent product faces. Once an agent can run in the background, the number of model calls is no longer tied to the number of chat messages the user typed. A single delegated job might include planning, retrieval, file reading, tool calls, draft generation, revision, validation, and final packaging. If the product hides all of that behind a flat unlimited promise, either the vendor eats the cost, limits the product, or quietly degrades quality.

Usage-based pricing is not automatically bad. It can be the honest version of an agent product. The buyer’s job is to make it measurable. Cowork’s announced inputs — model use, context retrieval, tool calls, and runtime — are the four buckets to ask every vendor about.

What buyers should ask before enabling Cowork-style agents

Start with model choice. Which models can run the task? Which model is the default? Can admins restrict or allow cheaper models? If an open-source model is offered through a managed platform, is it optional, and can teams see when it was used?

Then ask about the data boundary. Microsoft’s March Cowork announcement says Cowork grounds work in emails, meetings, messages, files, and data through Work IQ, and that the system turns requests into plans, runs in the background, checks in for clarification, and supports approval before applying changes. That is useful only if the organization can see which sources were retrieved and which actions were proposed. For permission-heavy workflows, pair this buying question with Toolhalla’s agent write-permission UX checklist.

Next, ask for budgeting controls. Can admins cap credits by user, team, task type, or connector? Can they distinguish a light research task from a heavy multi-output job? Can finance see which teams are generating the highest agent runtime? Microsoft’s GA post says Cowork includes new cost management controls, but each organization still needs its own guardrails before broad rollout.

Finally, test the fallback path. If a cheaper model is good enough for routine analysis but not for legal, financial, or code-sensitive work, the platform needs routing rules. Toolhalla has covered the same pattern in LLM gateway decisions: OpenRouter vs LiteLLM vs Portkey is the infrastructure version of the same buyer question. The enterprise-agent version is: which model does this task deserve, and who pays when it is wrong?

Practical takeaway

Do not read the Copilot Cowork news as a simple DeepSeek headline. Read it as a pricing and architecture shift. Microsoft is telling customers that long-running agents have measurable costs across model use, retrieval, tool calls, and runtime. Axios is reporting that cheaper model options, possibly DeepSeek V4 or another open-source model, are part of Microsoft’s path to make that cost workable.

For teams evaluating Cowork, Claude-style agents, coding agents, or internal workflow agents, the checklist is now straightforward:

• Map which tasks are light, medium, and heavy before rollout.
• Require per-task cost visibility, not only per-seat pricing.
• Check which model is used for each task class.
• Verify the data boundary for company context and files.
• Keep human approval and read-back for write-capable workflows.
• Route lower-risk work to cheaper models only when quality and policy allow it.

Agent products will keep getting better, but the bill is becoming part of the product design. Copilot Cowork makes that visible.