Copilot shifts to Microsoft MAI models to cut OpenAI costs

Microsoft is replacing OpenAI and Anthropic models with its own MAI models across several Copilot products, including Excel and Outlook. The company says the MAI models are already running in apps such as GitHub Copilot, and Bloomberg reported MAI models are processing tens of thousands of requests per week in Excel and Outlook, though they still handle only a small fraction of total requests.

What is changing in Copilot and Office?

Microsoft is switching the backend models for Copilot and Office features from third parties to its in-house MAI models, rolling them into Excel, Outlook and GitHub Copilot and preparing a proprietary transcription model for Teams. The move is explicitly framed as a cost-cutting step: Mustafa Suleyman, Microsoft’s head of AI, said "We pay a lot of money to Anthropic—so our goal is to reduce and ultimately eliminate that cost."

The company unveiled seven new AI models at the Build conference, including MAI-Thinking 1, billed as a reasoning model. Microsoft also says the MAI models are trained on clean, commercially licensed data, but its technical paper shows the use of the Common Crawl dataset, a freely accessible web corpus whose legal status for training is unsettled.

How do MAI models perform compared with OpenAI and Anthropic alternatives?

MAI-Thinking 1 was positioned by Microsoft as a competitive reasoning and coding model, but independent benchmarks painted a different picture: Microsoft claimed Thinking-1 could match Sonnet 4.6 and Opus 4.6 on coding based on human evaluations, yet the benchmarks Microsoft released showed Thinking-1 trailing the competition from OpenAI and Anthropic by a wide margin and landing roughly on par with Deepseek V3.2.

In short, Microsoft’s internal human-evaluation claim is more optimistic than the public benchmark set. The source material places MAI-Thinking 1 below the leading OpenAI and Anthropic models on the objective tests Microsoft published, even as Microsoft pushes MAI as the default for cost and licensing reasons.

Why it matters

Microsoft stands to lower its AI bill by routing more Copilot traffic to MAI models, shifting OpenAI and Anthropic costs off its balance sheet. CEO Satya Nadella signaled that billing could move more toward usage-based pricing instead of flat-rate subscriptions, which opens a likely setup where cheaper MAI models are the default and third-party models become premium add-ons. That would keep headline subscription prices steady while passing the incremental cost of higher-performing third-party models to customers.

The quality trade-off matters for enterprises that pay for Office and Copilot features. MAI models are being presented as safer and commercially licensed, but the disclosed training mix includes Common Crawl. The combination of weaker benchmarks for MAI-Thinking 1 and the use of public web data undercuts a simple claim of superior data hygiene.

What to watch

Track three concrete signals: whether Microsoft expands MAI deployment beyond the current apps, how quickly MAI share of requests rises from the current "small fraction" mentioned in reporting, and whether Microsoft shifts Copilot billing toward the usage-based or add-on pricing Nadella hinted at. Another immediate marker is whether Teams ships the proprietary transcription model Microsoft expects to deliver and how it is priced compared with existing options.

Microsoft’s strategy ties product defaults to cost control. The near-term result will reveal whether large customers tolerate lower baseline model capability for a lower effective price, or whether demand for top-tier OpenAI and Anthropic models persists as a billable premium.