Microsoft has a new plan for AI: Use expensive frontier models only when needed
&imwidth=600&imheight=450&format=webp&quality=medium "Microsoft has a new plan for AI: Use expensive frontier models only when needed")
Microsoft copilot
Microsoft is expanding its home-grown MAI models across GitHub Copilot, Excel and Outlook. Satya Nadella says the strategy uses specialised, cheaper models for everyday tasks while keeping powerful OpenAI and Anthropic models available for demanding workloads.
Microsoft is changing the way it powers Copilot, and the strategy could have major implications for OpenAI, Anthropic and the economics of enterprise AI. CEO Satya Nadella says Microsoft is increasingly using its own MAI family of AI models across products including GitHub Copilot, Excel and Outlook. Instead of relying on the most powerful and expensive frontier model for every request, Microsoft wants to use the right model for each job. The goal is simple: maintain or improve AI performance while using fewer tokens and reducing the enormous computing cost of serving millions of users.
Microsoft's own AI is already taking on GPT and Claude
Microsoft says early results are promising. Its specialised MAI model running in Excel is performing on par with GPT-5.6 for common tasks while being cheaper to operate. In GitHub Copilot, MAI-Code-1-Flash has recorded an approximately 10% higher code acceptance rate than GPT-5.4 Mini and Claude Haiku 4.5 in Microsoft's testing. It also used around 10% fewer median tokens. Rather than training one enormous model to do everything, Microsoft trains specialised models inside environments based on the products in which they will actually operate. That means an Excel AI learns spreadsheet workflows, while a coding model learns how developers work inside GitHub Copilot.
OpenAI and Anthropic are not disappearing
The shift does not mean Microsoft is abandoning frontier models. Nadella said models from OpenAI and Anthropic will remain part of Microsoft's AI orchestration system. Powerful models can handle the hardest requests, while smaller and cheaper MAI models deal with high-volume everyday work. The model itself, Nadella argues, is only one part of an AI product. Memory, tools, context, user interactions and the surrounding software can be just as important.
Microsoft calls its continuous optimisation process "hill climbing." Instead of judging models only through standard benchmarks, Microsoft evaluates whether they successfully complete tasks customers actually care about.
Trending Stories
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
&im=FitAndFill=(700,400))
)
Why Microsoft's strategy matters
AI has introduced something unusual into mainstream software: a significant cost every time a user asks the system to perform work. For products serving hundreds of millions of people, even small reductions in inference costs can therefore become extremely valuable. Microsoft's answer is a multi-model system where expensive frontier intelligence is used only when necessary. The company says it is seeing promising early results in GitHub Copilot, Excel and Outlook, with the approach now expanding towards Copilot Chat and PowerPoint. Microsoft also plans to bring the same framework to businesses through Microsoft Foundry, allowing companies to optimise AI agents around their own data, workflows and evaluations. The bigger shift is clear: the next stage of the AI race may not simply be about who builds the world's smartest model. It may be about who can deliver enough intelligence for each task at the lowest possible cost.
