Summary
- Microsoft is evaluating DeepSeek V4 — or another open-source model — as a cheaper backend for Copilot Cowork, with a decision expected within weeks.
- The price gap is stark: Anthropic’s latest flagship costs roughly $50 per million tokens; DeepSeek V4 Pro runs at approximately $0.87 per million tokens — a 57x difference.
- Any DeepSeek deployment would be optional, fully hosted on Azure, and covered by Microsoft’s enterprise compliance and data residency controls.
- Copilot Cowork has switched to usage-based billing through Copilot Credits, following GitHub Copilot’s same move on June 1.
- Microsoft has reportedly already restricted Claude Fable 5 over data retention concerns, signaling friction with Anthropic beyond just cost.
The Number That Makes This Make Sense
Anthropic’s latest flagship model costs approximately $50 per million tokens, while DeepSeek V4 Pro comes in at around $0.87 per million tokens — a roughly 57x price gap. That’s not a rounding difference. That’s a structural problem for any company trying to run agentic workflows at scale, where an AI agent might chain dozens of model calls just to complete a single task like sorting through emails, drafting a report, or coordinating across Teams and Outlook.
Enterprise AI interaction costs have jumped 30x since 2023, and Goldman Sachs projects agentic workflows could drive token demand up 24x from current levels. The math stops working at some point. Microsoft appears to have decided that point is now.
What Copilot Cowork Actually Is
Copilot Cowork is Microsoft’s agentic AI feature within Microsoft 365, built in collaboration with Anthropic, that autonomously handles complex multi-step tasks across apps like Outlook, Teams, and Excel. It became generally available worldwide on June 16. Pricing requires a Microsoft 365 Copilot license at $30 per user per month, plus additional usage-based charges through Copilot Credits — with cost per task varying depending on which model is used, how much context is retrieved, how many tool calls are made, and total runtime.
That last part is the critical change. Flat-rate pricing made sense when AI usage was predictable. Agentic workflows aren’t predictable — they scale with task complexity, and complex tasks are exactly what businesses want to use them for.
“The 57x token price gap between Anthropic’s flagship and DeepSeek V4 Pro isn’t an abstraction — it’s the difference between AI agents being a competitive tool and an IT line item that scares CFOs.”
The DeepSeek Complication
Microsoft’s move puts the company on a collision course with both enterprise security teams and a White House that has spent months trying to wall off Chinese AI from American infrastructure. DeepSeek is a Chinese-developed model, and that origin matters politically regardless of where the inference runs. Microsoft says any DeepSeek deployment would be optional for customers and fully hosted on Azure, keeping customer data within Microsoft’s cloud and covered by Azure’s enterprise protections.
Whether that’s sufficient is a legitimate debate. Hosting the model weights on Azure addresses data residency. It doesn’t change where the model was trained, by whom, or on what data.
Not Just DeepSeek
Microsoft is expected to add other models — possibly Meta’s Llama 4 and Mistral’s latest — to Copilot Cowork over time. This suggests a multi-model strategy rather than a bet on any single alternative. Microsoft has also reportedly restricted Claude Fable 5 from Copilot Cowork. It was due to concerns related to data retention policies. This adds a layer of friction with Anthropic that goes beyond simple cost calculations.
GitHub Copilot made the same shift to consumption-based pricing on June 1, 2026 — signaling this is a deliberate, coordinated direction for Microsoft’s AI portfolio, not a one-off response to cost pressure.
