Microsoft is evaluating multiple open-weight models for its agentic "Copilot Cowork" feature. Starting with DeepSeek V4 and including GLM, MiniMax and Kimi, the company is moving toward making its foundation models "interchangeable."
Microsoft 365 Copilot · Agentic AI
Microsoft Wants Copilot Cowork's Brains to Be Swappable
Beyond DeepSeek, Microsoft is evaluating cheaper Chinese open-weight models — GLM, MiniMax and Kimi — to make the foundation model behind its task-automating agent interchangeable, separating the orchestration layer from the model itself.
~$25
Cost of one heavy Cowork task on Claude Opus 4.8
~$30
Per seat for a standard Copilot license (basic use)
1.04T
Parameters in Kimi K2.5 (MoE), with 256K context
SWE-bench Verified: the cost case for Chinese models
Open-weight challengers are approaching frontier scores — at lower cost.
~80%
GPT-5.2 /
70–80%
GLM · MiniMax
Takeaway: near-frontier task accuracy at a fraction of the price drives the swap-the-model strategy.
The interchangeable-model stack
Harness
Orchestration layer
Auto mode picks the optimal model per task
→
Interchangeable foundation models
Claude · GPT-5.5 · DeepSeek · GLM · MiniMax · Kimi · Cowork 1
Supported models & their roles
Claude Opus 4.8
Complex reasoning, high-difficulty tasks
Claude Sonnet 4.6
Everyday tasks, fast responses
GPT-5.5
Added for Frontier users
Cowork 1
Microsoft's own low-cost model (coming soon)
Upside
Multi-model choice improves task fit — Opus for complex work, Sonnet for everyday — and Claude integration boosts execution.
Concerns
Usage-based billing raises costs, and adopting Chinese models opens security, bias and reliability-review questions.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…