From Frontier APIs to Flexible Model Stacks
Microsoft’s shift toward cheaper open source AI models is a strategic move to contain enterprise AI costs while keeping Copilot-grade performance available at scale for everyday business workflows. Instead of tying Copilot Cowork to a single expensive proprietary model, Microsoft is designing the system so the orchestration layer stays stable while underlying models can be swapped based on task, latency, and price. That means some work can still go to frontier APIs from OpenAI or Anthropic, while other parts move to self-hosted open-weight models on Azure. Over time, smaller models may even run on local machines for lighter tasks. For enterprises, this promises more predictable AI model economics and better control over how different workloads map to different cost tiers, without losing the governance and security controls they expect from Microsoft 365.
DeepSeek V4 and the Search for a Cheaper Anthropic–OpenAI Alternative
The clearest sign of this shift is Microsoft’s evaluation of a fine-tuned, self-hosted DeepSeek V4 as a more affordable DeepSeek alternative to Anthropic and OpenAI models. Reports say DeepSeek V4 delivers strong performance at a fraction of the per‑token price of frontier models, making it attractive for long-running agentic workflows like email triage, meeting prep, and cross‑app reporting. Any DeepSeek deployment would run entirely on Azure, keeping customer data inside Microsoft’s cloud with standard compliance and residency controls. That structure helps address concerns about using an externally developed model while still benefiting from open-weight economics. If DeepSeek or similar open models can handle a large share of enterprise tasks, proprietary model providers will face sharper price pressure as companies push for better enterprise AI costs without accepting a big drop in quality.
Usage-Based Pricing and the New AI Model Economics
Microsoft’s model experiments arrive alongside a decisive pricing turn: AI agents are moving from subscription comfort to metered compute. GitHub Copilot now charges in AI Credits instead of a flat premium plan, and Copilot Cowork has followed with usage-based pricing tied to task complexity and token consumption. According to Startup Fortune, a GitHub user projected an USD 847 (approx. RM3,910) bill under the new system after previously paying USD 39 (approx. RM180) a month, highlighting how agentic AI no longer behaves like autocomplete. The old “pay per seat, use without limit” deal is ending because multi-step agents that plan tasks, open files, call tools, and self-check outputs are expensive to run. Enterprises now need cost controls, model policies, and clear guidance on which tasks deserve high-end models and which should run on cheaper open source AI models.
Internal Pressure and the Rise of Scout-Style Autonomous Agents
This cost-aware pivot also reshapes incentives inside Microsoft. Copilot Cowork’s model-agnostic design means internal MAI teams must compete head-on with fast-moving open-model providers for real workloads. If Chinese and open-weight models continue to perform well in production, they become a permanent benchmark for Microsoft’s own research groups. In parallel, the company is rolling out more autonomous systems, such as the Microsoft Scout Autopilot agent, that can take on multi-step tasks with less human supervision. To keep those agents affordable, Microsoft needs a stack of models tuned not only for capability but also for price and latency. The result is a new kind of optimisation problem: pairing each step of an agentic workflow with the lowest-cost model that still meets quality and compliance requirements, instead of defaulting everything to a single frontier API.
A Tipping Point for Enterprise AI Costs
Taken together, open model evaluations, DeepSeek trials, and metered Copilot pricing signal that enterprises have reached a tipping point on AI spend. Agentic tools like Copilot Cowork are useful enough to generate heavy usage across Outlook, Word, Excel, and Teams, but too expensive to hide inside flat licenses. As Startup Fortune notes, the real lesson is that AI procurement now resembles cloud infrastructure: every agent is a worker with a meter attached, and every model choice carries a cost curve. Microsoft’s response—model-agnostic orchestration, optional open-weight deployments on Azure, and granular AI credits—shows where the market is heading. Buyers will demand more open source AI models as viable defaults, keep premium frontier APIs for high-stakes tasks, and insist on transparent AI model economics before rolling autonomous agents out across the business.
