The end of blind trust in third-party models
Microsoft has been one of the largest customers of external AI models so far. It invested 13 billion dollars in OpenAI and has long purchased models from Anthropic to power its Copilot ecosystem. But now, according to Bloomberg, it is starting to rely on itself.
Internal models labeled as MAI are already running in two of the most widely used office applications in the world — Excel and Outlook. They are replacing tasks there previously handled by models from OpenAI (GPT) and Anthropic (Claude). This is Microsoft's first operational departure from external AI suppliers within its flagship commercial software.
Why Microsoft is changing strategy
The reason is economic. Running AI features for hundreds of millions of Microsoft 365 users is not cheap. Every query that goes through the API from OpenAI or Anthropic costs money. When scaled to tens of millions of users daily, fractions of cents turn into hundreds of millions of dollars per year.
Mustafa Suleyman, head of Microsoft's AI division, has previously hinted that the company plans to reduce and eventually completely eliminate spending on models from Anthropic. Moreover, some promotional phases of the OpenAI partnership are starting to wind down, which frees Microsoft's hands to deploy its own solutions more aggressively.
According to Bloomberg, MAI models already handle tens of thousands of user queries per week. While this represents only a fraction of Copilot's total traffic so far, it is proof that the internal models are performant enough for real deployment in a production environment.
What MAI models can do and where they're headed next
MAI (Microsoft AI) is a family of proprietary models developed internally. Microsoft is silent about their architecture, but according to available information, they are large language models optimized for specific product scenarios — not general-purpose chatbots like ChatGPT or Claude.
Deployment is not limited to Excel and Outlook. Microsoft has also begun integrating its own AI into GitHub Copilot, its flagship product for assisted programming. In the coming months, it plans to deploy an internally developed speech-to-text model into the Microsoft Teams platform.
For context — GitHub Copilot has over 50 million users and has become one of the fastest-growing AI products in history. Switching to its own models at such a massive scale could save Microsoft hundreds of millions of dollars annually.
What this means for OpenAI and Anthropic
For OpenAI, this is unpleasant, though not surprising, news. Microsoft was not only its largest investor but also its largest customer. Losing even a partial volume of API calls from Microsoft 365 represents a significant revenue hit. OpenAI is preparing for this eventuality — it recently established direct relationships with enterprise customers and is expanding into AI agents and financial services.
For Anthropic, this is another signal that relying on Microsoft as a key customer was not a sustainable long-term strategy. Suleyman's statement about "eliminating spending on Anthropic" gives this departure a clear timeframe.
Interestingly, while Microsoft is backing away from OpenAI and Anthropic, Anthropic itself is testing chips from Microsoft as an alternative to Nvidia. Relationships between tech giants in the AI era are simply becoming more complex and less transparent.
Stock reacts positively
The market received the news with optimism. Microsoft's stock gained approximately 2% after the information was published. Investors see the cost savings on AI infrastructure as a step in the right direction — especially at a time when AI spending at large tech companies reaches tens of billions of dollars per year.
Microsoft has lost approximately 20% of its value over the past year, so any signals about improving unit economics are welcomed by investors.
Impact on Czech companies and users
For Czech companies and individuals using Microsoft 365 and Copilot, nothing changes in the short term — functionality remains the same, it's just newly powered by Microsoft's internal models instead of GPT or Claude. In the long run, however, this transition could mean lower prices for end customers, as Microsoft will no longer pay margins to external suppliers.
Copilot for Microsoft 365 is available in the Czech Republic as part of the Microsoft 365 Business Standard and Premium subscription for an additional fee of approximately 30 USD (roughly 700 CZK) per user per month. GitHub Copilot is available for free in the basic version, with paid plans starting at 10 USD per month. However, neither of these products is currently localized into Czech — the interface and communication are in English, or other world languages.
Does this mean Microsoft is completely ending its collaboration with OpenAI?
No. Microsoft remains the largest investor in OpenAI with 13 billion dollars and continues to use GPT models in the Azure OpenAI Service product. This is a gradual introduction of its own models where it makes economic sense — particularly for routine, high-frequency tasks in office applications. The strategic partnership with OpenAI continues, only its intensity in specific product scenarios is changing.
Are Microsoft's MAI models as powerful as GPT-5.5 or Claude Opus 4.8?
Microsoft has not published any benchmarks for its MAI models, so a direct comparison is not possible. However, it can be assumed that for specialized tasks in Excel and Outlook (data analysis, email summarization, formula generation) they are sufficiently performant — otherwise Microsoft would not deploy them in production. For more complex tasks such as advanced programming or creative writing, external models will likely continue to be used.
When will the change be noticeable for regular Microsoft 365 users in the Czech Republic?
The transition is already underway, but it is practically invisible to the end user — functionality remains the same. The only difference you might notice is a potential change in the quality or speed of responses for certain AI features in Excel and Outlook. Microsoft tests these transitions gradually and evaluates feedback.