Microsoft Eyes a DeepSeek V4 Engine to Cut the Cost of Copilot Cowork

Microsoft Weighs a DeepSeek V4 Model for a Cheaper Cowork Tier

Microsoft is evaluating a fine-tuned version of DeepSeek V4 to run a lower-cost tier of its Copilot Cowork enterprise agent. The deliberation surfaced on the same day the company moved Cowork to usage-based pricing and made the tool generally available worldwide, signaling that affordability has become a central concern as the product reaches a global audience.

The company has not formally committed to DeepSeek V4 as the cheaper alternative. It has, however, said a lower-cost model option would be confirmed and available within weeks, leaving little doubt that a budget engine is coming even if the specific model remains undecided. Microsoft is also preparing its own "Cowork 1" branded model, described as a secure option for everyday tasks at substantially lower cost, which points to a deliberate strategy of stacking multiple price-performance options inside a single product.

Why Soaring Token Costs Are Pushing Microsoft Toward a Budget Engine

The search for a cheaper engine is driven by how heavily some organizations lean on the tool. Charles Lamanna, Microsoft's executive vice president for Copilot, Agents, and Platform, said some enterprise users run hundreds of tasks per week through Cowork, pushing costs very high. At that volume, the expense of running an always-on agent stops being a rounding error and starts shaping whether a deployment is sustainable.

How Agentic Workflows Burn Through Compute

Cowork's cost profile differs sharply from a conventional chatbot exchange. Rather than returning a single reply, the agent plans multi-step workflows, retrieves context, calls tools, and iterates until a task is finished. Every loop in that process consumes model tokens and compute, so a single complex assignment can be far more demanding than a one-off question. Multiply that across hundreds of weekly tasks and the economics explain why Microsoft is hunting for a model that can do the same work for less.

The Models Running Copilot Cowork at General Availability

At launch, Cowork relies primarily on Anthropic models, including Opus 4.8 and Sonnet 4.6. OpenAI's GPT 5.5 is available in the product's Frontier preview tier, giving advanced users access to a higher-end engine. The potential addition of a DeepSeek-based option, alongside the forthcoming Cowork 1 model, would round out a lineup that spans premium frontier capability at the top and low-cost, everyday processing at the bottom.

This model diversity is the through-line of Microsoft's approach. By matching different engines to different price points and task types, the company is betting it can keep heavy users on board without absorbing the full cost of running premium models on every request.

Azure Hosting Is Built to Contain Security and Compliance Risk

The optional DeepSeek-based model would be fully hosted on Microsoft's Azure infrastructure. Customer data would remain inside the tenant boundary and stay covered by the existing enterprise security, compliance, and data-residency controls that organizations already depend on. Microsoft said it has fine-tuned the model and added safeguards, including adjustments aimed at reducing bias, positioning the option as something it controls rather than a raw third-party drop-in.

The arrangement mirrors the playbook Microsoft has used through Azure AI Foundry, where third-party models are offered under the company's own operational and contractual umbrella. That structure lets customers tap an outside model while keeping Microsoft as the accountable party for hosting, terms, and governance.

Why a Chinese-Developed Model Invites Extra Scrutiny

Even with Azure hosting and added safeguards, putting a Chinese-developed model inside an enterprise productivity tool is likely to draw scrutiny from regulated industries and government customers. For these buyers, provenance and data handling are not abstract concerns but procurement requirements, and the origin of the underlying model can become a sticking point regardless of where it runs or how it is secured.

Usage-Based Pricing Reshapes the Enterprise AI Deal

Cowork's shift away from flat-rate licensing marks a broader reckoning for agentic AI products. Instead of a predictable per-seat fee, Microsoft is charging in "Copilot Credits," with costs driven by model use, context retrieval, tool calls, and runtime. The pricing now tracks the actual work the agent performs, which aligns spending with consumption but also makes budgeting less predictable for organizations with uneven usage.

Copilot Credits and Administrative Spending Controls

To keep that variability in check, Microsoft has built in guardrails. The tool is off by default, and administrators can set spending limits at the tenant, group, and user levels. That gives IT teams a way to cap exposure before usage runs away, which matters in a model where a handful of power users can generate outsized costs.

Microsoft has claimed internally that Cowork is 30 to 40 percent cheaper than Anthropic's Claude Cowork paired with a Microsoft 365 connector, while acknowledging that costs vary by configuration. Layering an even cheaper model tier on top of that comparison underscores the central tension: the economics of always-on AI agents remain a barrier to broad enterprise adoption, and Microsoft is treating model diversity as its path to making the math work.

Worldwide Availability, Partner Plugins, and Web Browsing

The pricing changes arrived alongside Cowork's general availability worldwide. Microsoft introduced nine partner plugins from companies including Monday.com, Miro, and Moody's, with Adobe, Atlassian, and others slated to follow. Frontier users can also direct Cowork to browse the web through a local Edge browser, extending the agent's reach beyond the data and tools already connected to it and broadening the range of tasks it can complete.