Microsoft replaces OpenAI's technology with its own AI in certain applications.

      TL;DR: Microsoft has started substituting OpenAI and Anthropic models with its own in-house MAI models in certain application features where cost or data residency is advantageous, according to Bloomberg. This transition is gradual, as OpenAI and Anthropic still manage the majority of Copilot's traffic, but it represents a consistent effort by Microsoft to lessen its reliance on these partners after investing years and billions to build that relationship. The change was made possible by a 2025 renegotiation that allowed Microsoft to create competing models.

      As reported by Bloomberg, Microsoft is replacing OpenAI and Anthropic models with its own AI for select product features. This update directs certain tasks to Microsoft’s internal MAI models when they offer cost or data residency benefits.

      The change is gradual rather than a complete overhaul. OpenAI and Anthropic continue to manage most production traffic within Copilot, while MAI models step in where they are economically viable.

      At its Build conference, Microsoft introduced seven MAI models, which include its first reasoning model as well as systems for image, voice, and transcription. It has been testing MAI-Transcribe-1 across Teams and Copilot and integrating MAI-Image-2 into Bing and PowerPoint.

      The focus is on efficiency, as Microsoft can utilize its Azure infrastructure for these models, avoiding third-party costs. It claims one MAI model, specifically optimized for consulting firm McKinsey, achieved ten times the cost efficiency of OpenAI’s GPT-5.5.

      Despite this shift, it doesn’t indicate a complete break. Microsoft launched the MAI series to directly compete with OpenAI while still retaining access to its partners’ models for their specific strengths.

      For years, Microsoft’s initial contract with OpenAI restricted it from independently developing advanced AI. This changed after the two firms renegotiated, allowing Microsoft to build competitive models while still holding a license to OpenAI’s technology through 2032.

      The agreement also permitted OpenAI to partner with rivals, enabling AWS to offer its models. Microsoft’s strategy has involved a three-pronged approach, maintaining a significant stake in OpenAI, incorporating Anthropic’s Claude into Copilot, and now deploying its own MAI models.

      The strategic reasoning behind this is something that Satya Nadella has hinted at, expressing concerns about Microsoft becoming “the next IBM” if it relied too heavily on one partner. By owning its models instead of merely renting them, Microsoft aims to avoid this outcome.

      Microsoft doesn’t require MAI to dominate every performance chart, as its access to hundreds of millions of Office and Teams users provides substantial leverage. Moving even a portion of that traffic to its own models generates significant revenue.

      However, this extensive reach comes with a persistent challenge, as Microsoft delves deeper into AI while the issue of monetizing these developments remains unresolved. More affordable in-house models improve profit margins regardless of whether users are charged.

      The costs associated with external AI are considerable, and every feature that Microsoft brings in-house reduces these expenses. The partnership that once characterized Microsoft’s AI narrative is being gradually adjusted, one model substitution at a time.