Microsoft explores DeepSeek V4 for Copilot Cowork to curb 'tokenmaxxing' costs

According to Axios, Microsoft is weighing a fine-tuned version of China's DeepSeek V4 to power Copilot Cowork, its newly general-available agentic productivity product. The motivation is blunt economics: long-running agents that chew through tokens on complex, multi-step tasks have driven compute costs to uncomfortable levels—what the community has dubbed 'tokenmaxxing'—and DeepSeek's efficiency-first architecture offers dramatically cheaper inference than the OpenAI and Anthropic models currently behind Copilot.

The cost problem is real and user-facing. On Reddit, a single agentic Copilot coding session reportedly burned $30–$40 in credits—three to four times a Pro subscriber's entire monthly allotment—fueling outrage that Microsoft is 'building features that encourage high token consumption, then penalizing users.' That backlash, combined with Microsoft's new consumption-based Copilot Credits billing, makes a cheaper underlying model strategically attractive.

Mechanically, Microsoft would fine-tune an open-source DeepSeek V4 (or similar) model and tailor it to specific Cowork workloads rather than route everything through premium frontier models—reserving expensive models for tasks that genuinely need them. Satya Nadella has already touted Copilot Cowork's GA 'with multi-model support,' signaling the architecture is built to swap models per task.

The obvious complication is geopolitics. Adopting a Chinese model—even an open-weights one running on Microsoft's own infrastructure—'probably to Trump's chagrin,' as Gizmodo put it, invites scrutiny precisely as Washington restricts Anthropic's models and weighs (then declines) blacklisting DeepSeek. The episode crystallizes 2026's central tension: Chinese open models offer compute savings too large to ignore, but using them carries political risk for US incumbents. Watch whether Microsoft proceeds, and how it frames data-handling and security if it does.