Microsoft Eyes DeepSeek V4 for Copilot to Cut Costs
Microsoft may swap OpenAI for DeepSeek V4 in Copilot to slash expenses—what it means for developers and model pricing.
The story of the day
Microsoft is evaluating DeepSeek’s V4 as a cheaper alternative to power Copilot Cowork, according to Axios. The move signals a shift from OpenAI’s models, likely driven by cost pressures—DeepSeek’s open-source models undercut proprietary pricing while matching performance on many tasks. For developers, this could mean lower API costs if Microsoft passes savings downstream, but also potential fragmentation in model compatibility. The bigger question: if Microsoft adopts a Chinese-origin model (DeepSeek is Beijing-based), how will export controls and compliance teams react?
New tools & launches
Wolfram Language 15 — Computational language adding native AI agent support for symbolic and numerical workflows. [Product Hunt]
Android 17 — Google rebrands Android as an "intelligence system," baking AI deeper into the OS. [Product Hunt]
BrainFlow — Turns voice memos into structured notes, open-source. [Product Hunt]
Henji — Trains AI to mimic your writing style for personalized replies. [Product Hunt]
Models & research
GLM-5.2 — Hugging Face’s latest open-weight model targets long-horizon tasks (e.g., multi-step reasoning, agentic workflows). No benchmarks yet, but positioning suggests competition with closed-source agents. [HuggingFace Blog]
Industry pulse
Microsoft’s reported pivot to DeepSeek V4 underscores the tension between cost and control. Axios notes the move is still exploratory, but the signal is clear: even hyperscalers are shopping for cheaper inference. OpenAI’s margin erosion may accelerate if Microsoft’s experiment succeeds. [The Information AI]
Stat of the day
DeepSeek’s V4 costs ~70% less than comparable proprietary models for similar performance on coding tasks, per third-party evaluations. [Axios]
Get the full analysis — subscribe to Gradient Pro.
