Microsoft Is Quietly Swapping OpenAI and Anthropic Out of Copilot to Cut Costs

Microsoft is replacing the AI models from OpenAI and Anthropic inside several Copilot products with its own in-house models, Bloomberg reports. The company’s MAI models are already handling tens of thousands of requests per week in Excel and Outlook — apps that previously leaned heavily on GPT and Claude models.

The in-house share is still a small fraction of total traffic, but the direction is explicit. Microsoft’s AI chief Mustafa Suleyman said it plainly in June: “We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.” The MAI models are also available in GitHub Copilot, and an in-house transcription model is expected in Teams soon.

Here’s the uncomfortable part: the swap isn’t happening because the in-house models are better. At its Build conference, Microsoft presented seven new MAI models, including MAI-Thinking 1, its first reasoning model — one designed to work through problems step by step before answering. Microsoft claimed it matched Anthropic’s Sonnet 4.6 and Opus 4.6 in coding based on human evaluations, but its own published benchmarks showed Thinking-1 trailing OpenAI and Anthropic by a wide margin, landing roughly on par with DeepSeek V3.2 — a capable open model, but not the frontier.

What’s behind it is simple economics. Running frontier models from other companies at Microsoft’s scale costs enormous sums, and those fees go to competitors. One likely future setup, per the reporting: cheaper MAI models become the default, while OpenAI or Anthropic models turn into premium add-ons at extra cost. Microsoft CEO Satya Nadella has also hinted that AI billing could shift toward usage-based pricing rather than flat subscriptions. One more wrinkle: Microsoft has marketed MAI models as trained on clean, commercially licensed data — yet its own technical paper lists Common Crawl, a scrape of the open web whose legal status for AI training is unresolved. That’s industry standard practice, but it sits awkwardly next to the marketing.

What this means for you: If you use Copilot in Office, watch quality, not announcements. The model behind a feature can change without any visible notice — if Copilot in Excel or Outlook suddenly feels less sharp, this may be why. For most people nothing breaks today; the models handle routine tasks fine. But it’s a useful reminder that “Copilot” is a brand, not a specific AI — and what’s underneath is increasingly chosen by cost, not capability.

Sources

• Bloomberg: Microsoft replaces OpenAI, Anthropic with own AI in some apps
• The Decoder: Copilot goes cheap as Microsoft phases out OpenAI and Anthropic models