bytevyte
bytevyte
Language
EN DE ES FR PT IT
Follow
ai-beats —

Microsoft MAI Models Cut Copilot Costs in Excel, Outlook

Microsoft MAI models

Microsoft has begun routing tens of thousands of weekly Copilot prompts inside Excel and Outlook to its own Microsoft MAI models, marking the first disclosed production-scale shift away from the OpenAI and Anthropic systems that have powered the company's AI features since launch. The change, confirmed by Bloomberg in early July 2026, is a deliberate effort by Microsoft to curb its fast-growing AI infrastructure costs by reducing reliance on third-party model providers.

The Microsoft MAI models are now handling a meaningful but still small fraction of total Copilot requests in the two Office applications, with volume measured in tens of thousands of prompts per week. Both Excel and Outlook previously depended more heavily on models from OpenAI and Anthropic to deliver their AI-assisted features. Microsoft has also deployed MAI models in GitHub Copilot, and the company expects to ship a proprietary transcription model in Teams in the near future.

The Cost Calculus Behind the Switch

Microsoft AI CEO Mustafa Suleyman has been direct about the company's motivation. In June 2026, he stated that Microsoft pays substantial fees to Anthropic and that the company's objective is to reduce and eventually eliminate that spending. For a hyperscaler running AI at the scale of Microsoft 365, the per-token cost of calling external model APIs adds up quickly across hundreds of millions of users. Shifting workloads to in-house models cuts that marginal cost to near zero after the initial training and inference infrastructure investment.

This cost-driven logic applies across the company's AI portfolio. While the current volume routed to Microsoft MAI models remains a small percentage of total Copilot traffic, the direction of travel is unmistakable. Microsoft is building toward a future where its own models serve as the default inference layer for its consumer and productivity AI features, with third-party models reserved for tasks where internal alternatives cannot meet quality requirements.

What Microsoft MAI Models Mean for Enterprise Users

The strategic shift introduces a tension that enterprise customers must weigh. Microsoft unveiled seven new AI models at its Build conference, including MAI-Thinking 1, its first reasoning model. However, benchmark results showed MAI-Thinking 1 trailing competing offerings from OpenAI and Anthropic by a wide margin. For Copilot and Office subscribers, this could mean paying the same monthly subscription price for AI features that run on less capable underlying models.

Microsoft has not disclosed whether it plans to differentiate pricing based on the model serving each Copilot request. The current Copilot subscription tiers for Microsoft 365 remain unchanged. If a significant portion of Copilot queries shifts to lower-performing MAI models without a corresponding price adjustment, enterprise buyers could face a de facto degradation in the value they receive for their AI spend. The quality gap matters most for complex analytical queries in Excel or nuanced drafting and summarization tasks in Outlook, precisely the use cases where corporate power users depend on Copilot.

That said, many routine Copilot requests, such as simple formula suggestions in Excel or short email replies in Outlook, may not require frontier-level reasoning capability. For those tasks, an adequate in-house model delivers the same user experience at a fraction of the cost. The question for Microsoft is how cleanly it can segment workloads: routing simple queries to MAI while preserving access to OpenAI and Anthropic models for complex ones, without introducing latency or reliability issues that frustrate users.

Structural Implications for the AI Ecosystem

The move carries consequences that extend well beyond Microsoft's own products. OpenAI and Anthropic have relied heavily on Microsoft as both a customer and a distribution partner. Microsoft's Azure cloud has been the exclusive compute provider for OpenAI, while Anthropic has also maintained a close commercial relationship with the company. A gradual reduction in model procurement from these partners is a revenue risk that both frontier labs will need to address.

For OpenAI, losing a portion of its largest enterprise channel means the company must accelerate its own direct enterprise sales efforts and diversify its customer base beyond the Microsoft ecosystem. Anthropic faces a similar challenge: Microsoft AI CEO Suleyman specifically named Anthropic as a cost line the company aims to eliminate entirely over time. Both labs have been building out their own enterprise go-to-market motions and competing for Azure-adjacent customers, but replacing the volume that Microsoft commands within its own products is a different scale of challenge.

The timing is also notable. This production shift comes at a moment when the broader AI industry is grappling with the economics of inference at scale. Frontier models are expensive to run, and the hyperscalers that deploy them at massive volumes are acutely sensitive to margin erosion. Microsoft's decision to build in-house alternatives mirrors a pattern already visible at Amazon with its Titan and Nova model families and at Google with Gemini. The hyperscalers are converging on a hybrid strategy: they offer third-party models through their cloud platforms as a revenue service while deploying their own models in their first-party products to control costs.

What This Means for Decision-Makers

For enterprise IT leaders and technology buyers, the Microsoft MAI models shift introduces a new layer of evaluation when assessing Copilot investments. The standard advice to audit AI feature usage and measure actual productivity gains becomes more urgent when the underlying model quality may shift without explicit notice. Companies running pilots or rolling out Copilot across their workforce should baseline current performance and satisfaction levels now, before a larger portion of queries migrates to Microsoft's in-house models.

From a vendor strategy standpoint, organizations that depend on frontier-model capabilities embedded in their productivity tools may need to supplement Copilot with direct API access to OpenAI, Anthropic, or other providers for high-stakes tasks. This creates a two-tier AI architecture within the enterprise: a built-in, cost-optimized layer for routine work, and a premium, best-in-class layer for complex or sensitive analysis. Budgeting for both tiers requires explicit planning.

For investors and industry analysts, the Microsoft MAI models rollout signals that the hyperscaler-AI lab relationship is entering a new phase. The partnership model that characterized 2023-2025, where cloud providers invested billions in frontier labs and distributed their models as a service, is giving way to a more competitive dynamic. The biggest customers of frontier AI models are also the best positioned to build their own alternatives, and they are acting on that capability. This structural shift will put pressure on independent AI labs to demonstrate differentiated value that cannot be replicated by the in-house efforts of their largest distribution partners.

Why this matters

Microsoft's decision to route Copilot traffic through its own MAI models is not an isolated cost-cutting measure. It is a strategic signal that the economics of enterprise AI favor the vertically integrated provider. When the company that controls the operating system, the productivity suite, the cloud infrastructure, and the AI model can serve its billion-user base at near-zero marginal inference cost, independent model providers face an existential competitive disadvantage. The next phase of the AI industry will be defined not by who builds the smartest model in isolation, but by who can deliver adequate intelligence at the lowest total cost across the widest distribution footprint.

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team.

Share