← All posts
standalone

Microsoft Copilot Swapping OpenAI and Anthropic Models for Its Own MAI

Microsoft is quietly replacing OpenAI and Anthropic models in Copilot with its own MAI models to cut costs. Here's what it means for developers using Copilot.

M
Mainthreaded ·2 min read ·1 source
#microsoft-copilot#mai-models#azure-openai#llm-apis#enterprise-ai

TL;DR

Microsoft is phasing out OpenAI and Anthropic models inside Copilot products like Excel and Outlook, replacing them with its own in-house MAI models. The stated goal is to eliminate external model costs entirely — which could affect the quality Copilot customers actually get.

What happened

Microsoft has begun routing Copilot queries through its own MAI models instead of using models from OpenAI or Anthropic. According to The Decoder, tens of thousands of queries per week are already running through these in-house models inside products like Excel and Outlook. Microsoft AI chief Mustafa Suleiman has stated that the company wants to “ultimately eliminate” the cost of paying for external models.

This is a meaningful strategic shift. Microsoft has invested billions into OpenAI and has been a marquee customer of Anthropic’s API as well. Quietly swapping those out for proprietary models in production workloads suggests the MAI program is further along than most outsiders realized.

Why it matters

Let’s be direct about what’s happening here: Microsoft is prioritizing margin over model quality, at least in the products where most enterprise users actually touch AI. Copilot for Microsoft 365 costs $30 per user per month. If the underlying model gets swapped from GPT-4-class performance to something cheaper and less capable, paying customers may notice — even if Microsoft doesn’t announce the change.

For developers building on top of Microsoft’s ecosystem (think Azure OpenAI Service, Copilot extensions, or M365 plugins), this is worth watching closely. The Azure OpenAI Service presumably still serves actual OpenAI models, since that’s the whole product promise. But if Microsoft’s internal apps increasingly run MAI instead, it signals that MAI is being battle-tested at scale. Depending on how good MAI actually is, this is either a smart vertical integration play or a quiet downgrade dressed up as cost efficiency. The sources don’t give us benchmark data to judge which, so skepticism is warranted.

What to watch

  • Whether Microsoft publicly discloses which model is handling which Copilot tasks — if they stay quiet, that itself is a signal.
  • Any enterprise Copilot customers reporting quality regressions in Excel or Outlook AI features, which would suggest MAI isn’t quite at parity yet.

Sources