Microsoft Shifts Copilot Cowork to Usage-Based Billing

Microsoft is changing how it charges for Copilot Cowork, moving the agentic assistant from a flat monthly fee to usage-based billing. The reason, in the words of Copilot chief Charles Lamanna, is that a flat rate simply doesn't hold up against the heaviest users. He pointed to people 'who do hundreds of tasks a week,' whose consumption pushes costs up faster than a fixed subscription can absorb. Alongside the pricing shift, Microsoft is weighing a cheaper engine under the hood: a self-hosted, customized version of DeepSeek V4 as an optional model, with a final decision said to be weeks away.

A bit of context on what Cowork is. It's Microsoft's reasoning-heavy, agentic layer of Copilot — built on Anthropic's Claude technology — that works across Outlook, Teams and Excel to carry out multi-step tasks rather than just answer one-off prompts. That agentic design is also why it burns through tokens quickly: doing real work, step after step, costs far more compute than a single chat reply. The proposed DeepSeek option would be fully hosted on Azure and include bias-mitigation safeguards, Microsoft says — framing clearly aimed at heading off the obvious questions about using a Chinese-developed model.

Part of a wider shift away from flat fees

This isn't a one-off. Microsoft already moved GitHub Copilot to token-based billing earlier in June 2026, and CEO Satya Nadella has been openly describing AI as a 'consumption business' — one where companies pick and tune different models for different jobs and costs, and where what you pay tracks how much you actually use. The Cowork change is the same logic applied to the office productivity side. The DeepSeek angle fits too: if usage is the cost driver, then offering a cheaper model for routine tasks is a natural lever to keep the bill — and Microsoft's own margins — under control.

Why this matters for you

The quiet headline here is that the all-you-can-eat era of AI pricing is ending. For a while, a flat subscription gave you effectively unlimited use, because providers were buying market share and betting most people wouldn't push hard. Agents broke that bet — they can grind through hundreds of tasks without getting tired — so the bill is moving to match the work. My take: this is healthier in the long run, because flat pricing for unlimited agent runs was never going to last, and pretending otherwise just delays the reckoning. But it does change how you should think about these tools. The cost of an AI assistant is becoming a variable line item, like electricity, not a fixed one like a streaming plan — and the people who notice that early will plan around it instead of being surprised by a spike.