Microsoft Begins Replacing OpenAI and Anthropic Models With In-House MAI AI in Excel and Outlook
- Microsoft is replacing OpenAI and Anthropic models with its own MAI models in Excel and Outlook, with tens of thousands of prompts already processed weekly [1]
- AI chief Mustafa Suleyman said Microsoft's goal is to "reduce and ultimately eliminate" the cost of Anthropic models, which Bloomberg reported cost roughly $500 million per year [2]
- Seven MAI models were unveiled at Build 2026, including MAI-Thinking 1, Microsoft's first reasoning model [3]
- GitHub Copilot and Teams transcription are next in line for the MAI model swap [1]
- MSFT shares rose 0.5% on the news to $388.84, though the stock remains down roughly 20% year-to-date [4]
Microsoft has begun swapping out AI models from OpenAI and Anthropic for its own internally built MAI models inside Excel and Outlook, Bloomberg reported on July 7 [1]. Tens of thousands of Copilot prompts per week in the two widely used Office applications are now handled by Microsoft's proprietary models rather than third-party systems, marking the first concrete step in a broader push to cut the company's AI inference costs.
The move reflects an explicit cost-cutting strategy articulated by Mustafa Suleyman, who leads Microsoft's AI division. "We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost," Suleyman said, according to Bloomberg [1]. Microsoft was spending approximately $500 million annually on Anthropic's models as of January 2026 [2].
The transition does not represent a full break from OpenAI or Anthropic. Both companies' models continue to handle the majority of Copilot traffic across Microsoft's product suite. But the shift signals a strategic pivot for a company that invested $13 billion in OpenAI and built its entire AI product strategy around the partnership — a relationship it renegotiated in 2025 to gain more freedom to develop competing models [1].
The MAI Models
Microsoft unveiled seven MAI models at its Build 2026 developer conference, spanning reasoning, coding, image generation, speech, and transcription [3]. The flagship, MAI-Thinking 1, is the company's first reasoning model. Microsoft claimed it matches Anthropic's Sonnet 4.6 and Opus 4.6 in coding tasks based on human evaluations, though independent benchmarks suggest the MAI family trails behind OpenAI and Anthropic's latest frontier models on broader tasks [3].
Suleyman has argued that Microsoft's approach is economically rational: the company can wait three to six months for the AI frontier to advance and then build a more efficient, specialized enterprise model rather than renting the newest external models for every interaction [3]. The first MAI models for speech-to-text, voice, and image generation launched in April 2026 [3].
Which Products Are Affected
Excel and Outlook are the first major Office applications to see the model swap, with tens of thousands of requests routed through MAI models weekly [1]. GitHub Copilot, Microsoft's AI-assisted coding tool, also has MAI models available for use [1].
Suleyman indicated that a Microsoft-built transcription model will begin powering Teams videoconferencing in the coming months, extending the MAI rollout beyond traditional Office productivity apps [1]. Microsoft declined to comment directly on the Bloomberg report [2].
Implications for OpenAI and Anthropic
The shift carries significant commercial consequences for Microsoft's AI partners. OpenAI, in which Microsoft has invested $13 billion, relies heavily on enterprise distribution through Microsoft's products to generate revenue. Anthropic, which counts Microsoft as a major customer, faces the prospect of losing a client that was spending an estimated $500 million a year on its models [2].
The move comes after Microsoft renegotiated its OpenAI partnership in 2025, gaining the contractual flexibility to build and deploy competing models [1]. That renegotiation now appears to have been a precursor to the current strategy of vertical integration — building in-house what it previously outsourced.
A Microsoft spokesperson had previously stated the company uses "a mix of models, which includes models from OpenAI as part of our partnership, as well as Microsoft AI and open-source models" [2]. The current transition makes clear that the mix is tilting increasingly toward Microsoft's own technology.
Market Reaction
Microsoft shares rose 0.5% to $388.84 on July 8, giving the company a market capitalization of approximately $2.89 trillion [4]. The stock remains down roughly 20% year-to-date from its 52-week high of $555.45, pressured by broader concerns about AI monetization and capital expenditure levels [4].
Analysts have generally viewed in-house model development as margin-positive for Microsoft, since inference costs for proprietary models are substantially lower than licensing fees paid to third-party providers. The question is whether MAI models can deliver comparable quality — a trade-off that Copilot customers will ultimately judge [3].
What's Next
Microsoft's MAI roadmap extends well beyond Excel and Outlook. The Teams transcription rollout is expected in the coming months, and the company's seven-model portfolio from Build 2026 suggests ambitions across the full AI stack — from reasoning and coding to multimodal generation [3].
The broader question is how fast Microsoft accelerates the transition. The current volume — tens of thousands of prompts per week — remains a small fraction of total Copilot usage. But Suleyman's stated goal of eliminating external model costs suggests the pace will quicken, particularly as MAI models mature and close the performance gap with frontier competitors [1].
Further sources
The stories that matter, in one email. Free — unsubscribe anytime.