Microsoft has begun shifting tens of thousands of weekly AI prompts in Excel and Outlook from OpenAI and Anthropic models to its own internally built MAI systems, Bloomberg reported on July 7. The move marks the first disclosed production-scale migration of Microsoft 365 Copilot traffic away from third-party AI providers, driven by an explicit cost-reduction mandate from Microsoft AI CEO Mustafa Suleyman, who told Bloomberg in June that the company’s goal is to “reduce and ultimately eliminate” its spending on Anthropic.
Key Takeaways
- Tens of thousands of weekly AI prompts in Excel and Outlook now run on Microsoft’s MAI models instead of OpenAI or Anthropic systems, the first confirmed production shift of Copilot traffic to in-house AI.
- Microsoft AI CEO Mustafa Suleyman stated the company’s goal is to “reduce and ultimately eliminate” Anthropic costs, which Bloomberg reported at roughly $500 million annually as of January 2026.
- Seven MAI models were unveiled at Build 2026, covering reasoning, coding, image generation, speech, and transcription, led by MAI-Thinking-1 for complex enterprise tasks.
- Microsoft’s OpenAI partnership runs through a 2032 license, after which the company would need to negotiate new terms or pay full market rates — making MAI development a long-term hedge against pricing exposure.
- OpenAI and Anthropic still handle the majority of Copilot traffic; the MAI shift currently covers a small fraction of overall usage.
What Changed Inside Excel and Outlook?
The Copilot interface in both applications looks identical to users. The button sits in the same place, accepts the same prompts, and returns the same types of outputs — formula suggestions in Excel, email summaries and draft replies in Outlook. What changed is the model processing those requests behind the surface. Bloomberg, citing a person familiar with the work, reported that tasks previously routed through OpenAI and Anthropic models are now being completed by Microsoft’s proprietary MAI models. Office Watch confirmed that no user-facing announcement accompanied the switch; Microsoft declined to comment.
The routing targets high-volume, routine prompts — the kind of constrained, predictable tasks that dominate everyday productivity workflows. Excel formula assistance and Outlook email summarization do not require frontier-grade reasoning; they require fast, accurate responses at scale, making them natural candidates for cost-optimized in-house models. Frontier-grade tasks can still route to OpenAI or Anthropic where the workload demands it, creating what enterprise analysts describe as a multi-model portfolio operating transparently behind a single Copilot interface.
Windows News reported that the transition stems from an internal initiative code-named “Project Orchard,” which outlined a target of reducing AI operational costs by up to 40% and cutting dependency on third-party models for routine productivity scenarios. Excel and Outlook were selected as starting points because their prompt patterns are relatively constrained and benefit from deterministic optimizations that favor smaller, task-tuned models over general-purpose frontier systems.
What Are the MAI Models and How Do They Compare?
Microsoft unveiled seven MAI models at its Build 2026 developer conference in June, representing the company’s first comprehensive in-house model family. The lineup includes MAI-Thinking-1 for complex reasoning tasks, MAI-Code-1-Flash for software development, MAI-Image-2.5 for image creation, MAI-Voice-2 for speech generation, and MAI-Transcribe-1.5 for speech-to-text. MAI-Thinking-1, the flagship reasoning model, operates with 35 billion active parameters, a 256,000-token context window capable of processing roughly 600 pages of text in a single pass, and a sparse Mixture of Experts architecture that activates only the subset of parameters relevant to each request — a design that keeps per-call inference costs low.
Microsoft claimed at Build that MAI-Thinking-1 matches the coding capabilities of Anthropic’s Opus 4.6 at reduced cost, and that a McKinsey-tuned version outperformed OpenAI’s GPT-5.5 on cost efficiency by a factor of ten based on public pricing comparisons. Those claims carry an important caveat: The Decoder noted that Microsoft’s performance benchmarks were largely produced by evaluations Microsoft commissioned, and independent benchmark testing showed MAI-Thinking-1 trailing current frontier models from OpenAI and Anthropic on broader task categories, landing roughly on par with DeepSeek V3.2. The gap between Microsoft’s internal claims and independent benchmarks suggests MAI models are competitive for specific enterprise workloads but are not yet general-purpose frontier replacements.
MAI models are already available in GitHub Copilot, and Suleyman confirmed that a Microsoft-built transcription model will be deployed in Teams in the coming months. The trajectory is clear: Microsoft is building a full-stack model portfolio designed to handle the majority of commodity AI workloads across its product suite, reserving external models for tasks where the performance gap justifies the cost premium.
What Is Driving the Cost Pressure?
The financial logic behind the shift is straightforward. Microsoft consumes massive volumes of AI tokens across Copilot, Azure, and its consumer products. Bloomberg reported that Microsoft was spending roughly $500 million per year on Anthropic models as of January 2026. While the company gets OpenAI models at a significant discount through its long-standing partnership, that arrangement has a defined endpoint: the license runs through 2032 as a fixed date, after which Microsoft would need to negotiate new terms or pay full market rates.
Suleyman framed the urgency in blunt terms at Build 2026. SiliconANGLE reported that he described Anthropic as “extremely expensive” and said many companies are “urgently looking for alternatives.” The comment reflects a broader industry dynamic: as AI adoption scales from experimentation into production workloads, the economics of per-token pricing from external providers become a structural cost problem for platform companies processing billions of requests across hundreds of millions of users.
Microsoft’s total projected capital spending for 2026 reached $190 billion — a figure that far exceeded Wall Street expectations when it was disclosed in April. At the same time, Microsoft’s stock has declined roughly 30% over the past nine months, erasing approximately $1.2 trillion in market value. The combination of massive AI infrastructure investment and investor pressure on profitability creates a mandate to reduce variable costs wherever possible, and third-party model fees are among the most controllable expenses in the AI stack.
What Does the Shift Mean for OpenAI and Anthropic?
The partnership structures remain intact. Microsoft is not ending its relationships with either company. OpenAI and Anthropic still handle the majority of Copilot traffic, and enterprise customers on Azure can continue selecting GPT, Claude, or MAI models depending on their workload requirements. FourWeekMBA described the new architecture as one where Microsoft is “demoting” external providers from default infrastructure to premium option — a distinction that matters for the long-term revenue models of both AI labs.
The risk for OpenAI and Anthropic is not losing the headline partnership agreements. It is losing volume. High-volume, low-complexity inference — the workload type that scales API revenue — is precisely what MAI targets. That volume erosion does not show up as a broken partnership; it shows up as a revenue growth curve that flattens more slowly than underlying model adoption would suggest. Tech Times reported that Anthropic projected roughly $10.9 billion in Q2 2026 revenue with Microsoft as one of its largest enterprise customers, making Suleyman’s stated goal of eliminating Anthropic spending a material concern for the company’s financial trajectory.
Microsoft also began canceling most internal Claude Code licenses in mid-May 2026, according to multiple reports, signaling that the shift extends beyond Copilot into developer tooling as well.
What Should Enterprise Customers Watch For?
The immediate user experience impact is negligible — Copilot surfaces remain unchanged, and Microsoft is managing the model routing transparently. Let’s Data Science recommended that enterprise AI teams treat Copilot as a model portfolio where cost, latency, governance, and vendor dependency can shift by task without notification. The more consequential question is whether Microsoft will document which workloads use MAI models, expose tenant-level controls, or adjust pricing and performance guarantees around specific Copilot features. For organizations operating under regulated workflows where model provenance and auditability matter, the absence of disclosure around which model is processing a given request creates a governance gap that IT administrators will need to address proactively.
The multi-model routing architecture — where Azure and Copilot dynamically assign each task to the optimal available model — is itself becoming Microsoft’s competitive moat. Enterprise customers do not need to pick a model; they buy the platform and the platform optimizes cost and quality automatically. That abstraction layer locks organizations deeper into Microsoft’s ecosystem than any single model partnership could.
Microsoft routing production AI workloads to its own models inside the two most widely deployed enterprise applications on earth is not an optimization — it is the beginning of a structural renegotiation of how value flows between platform companies and the AI labs that built the models they depend on.
FAQs
What are Microsoft’s MAI models?
MAI is Microsoft’s in-house family of AI models unveiled at Build 2026. The lineup includes seven models covering reasoning (MAI-Thinking-1), coding (MAI-Code-1-Flash), image generation (MAI-Image-2.5), speech (MAI-Voice-2), and transcription (MAI-Transcribe-1.5). The models are designed to handle enterprise workloads at lower cost than external providers.
Will Excel and Outlook Copilot features change for users?
The user-facing experience remains the same. The Copilot button, interface, and functionality are unchanged. The difference is in the model processing requests behind the interface — some prompts now route to Microsoft’s MAI models instead of OpenAI or Anthropic systems. Microsoft has not provided a user-facing toggle or disclosure indicating which model is handling a given request.
Is Microsoft ending its partnership with OpenAI?
Microsoft is not ending its partnership with OpenAI. The license agreement runs through 2032, and OpenAI models continue to handle the majority of Copilot traffic. The MAI shift targets routine, high-volume workloads where Microsoft can achieve comparable output quality at lower cost, while reserving external models for tasks that require frontier-level performance.
How much does Microsoft spend on external AI models?
Bloomberg reported that Microsoft was spending roughly $500 million per year on Anthropic models as of January 2026. The company also pays for OpenAI models, though at a discounted rate through its partnership. Microsoft AI CEO Mustafa Suleyman has stated the goal is to “reduce and ultimately eliminate” Anthropic spending specifically.
Which Microsoft products will use MAI models next?
MAI models are already live in GitHub Copilot and now in Excel and Outlook. Microsoft confirmed that a MAI-based transcription model is coming to Teams in the coming months. Facebook and Messenger integrations for other MAI capabilities, as well as PowerPoint support, have also been referenced in Build 2026 materials.
