bloomberg.com web signal

Microsoft swaps OpenAI image models out of PowerPoint and Bing

TL;DR

  • Microsoft is replacing OpenAI's image-generating models with its own MAI technology in PowerPoint and Bing, per Bloomberg.
  • In PowerPoint, Microsoft says its image model cut costs 84% compared with GPT-Image-2, alongside quality and retention gains.
  • The image swap extends an earlier July shift in Excel and Outlook, where MAI models began displacing OpenAI and Anthropic on routine prompts.

Microsoft is pulling OpenAI's image models out of two of its most visible consumer surfaces, PowerPoint and Bing, and replacing them with its own MAI-Image technology. Bloomberg reports that Mustafa Suleyman, who runs Microsoft AI, framed the swap in blunt operator terms: the in-house model is "faster, it's cheaper, it's higher quality, it drives better retention."

The number that makes the strategic logic obvious is the PowerPoint one. Microsoft's own image model, according to the reporting, cut costs 84% compared with GPT-Image-2 on that surface. For a product used inside enterprises at the volumes PowerPoint sees, an 84% unit-cost reduction on a feature that used to be metered through a partner's API changes what Copilot's gross margin can look like at scale, and it changes it without a price change the customer ever notices.

Read alongside the earlier July move to route Excel and Outlook prompts through MAI instead of OpenAI and Anthropic, this is the same pattern extending into image. High-volume, routine tasks get pulled onto Microsoft's own stack; the frontier stuff, presumably, still routes out. That is not a break with OpenAI, but it is the part of the relationship where Microsoft controls the meter.

The honest caveat is that most of the quality and retention claims here come from Suleyman himself, so take the specifics as reported rather than settled. What the reporting does not give you is how much of Bing image traffic is actually cut over versus still dual-routing, what the baseline for that "better retention" number is, or whether enterprise admins who care about which model touches their tenant get a switch. The 84% figure is compared to a specific model, not to whatever OpenAI ships next.

The forward-looking part is who this arms. Microsoft AI now has a demonstrable margin lever on Copilot that scales with usage, and MAI-Image becomes a more credible Foundry offering for developers who want a cheaper alternative without wiring in a competitor's brand. Adobe, Canva and Google get a live experiment in whether an in-house model can hold consumer-facing image quality at Bing scale. And OpenAI gets a preview of what its most important customer looks like when it stops needing to buy quite so much.