Microsoft’s MAI Models at Build 2026: A Fact-Checked Look at the New AI Family, Cost Savings, and What It Means for OpenAI

At Build 2026 in June, Microsoft announced a family of seven new in-house MAI models covering reasoning, coding, image generation, transcription, and voice ^M. The headline model is MAI-Thinking-1, a 35-billion-parameter reasoning model with a 256K context window, built from scratch on commercially licensed data ^B. For developers, Microsoft also released MAI-Code-1-Flash, a small-tier coding model tuned specifically for GitHub Copilot ^D. The full lineup includes MAI-Image-2.5 and its Flash variant, MAI-Transcribe-1.5, MAI-Voice-2 and its Flash variant, and MAI-Code-1-Flash ^WY.

Less than a month later, on July 7, 2026, Bloomberg reported that Microsoft had begun deploying these MAI models inside Excel and Outlook — replacing OpenAI and Anthropic models for tens of thousands of weekly AI prompts ^BF. The stated motivation: cost reduction ^BF. These prompts had previously relied on third-party models; now Microsoft is quietly routing them to its own models through what the company calls model orchestration on its Azure AI Foundry platform ^B.

Pricing Advantages: 20–60% Cheaper

A June 2026 MAI pricing comparison, citing Azure AI Foundry data, reports that MAI models are 20–60% cheaper than comparable third-party models ^M. Specific examples from that analysis:

Model	Price per 1K input tokens	Comparable Third-Party Model	Third-Party Price
MAI-Thinking-1	$0.03	OpenAI o3 equivalent	$0.10 or more
MAI-Code-1-Flash	$0.015	Claude Haiku 4.5	$0.025

On those numbers, MAI-Thinking-1 is roughly 70% cheaper than the cited OpenAI o3-equivalent, while MAI-Code-1-Flash is 40% cheaper than Claude Haiku 4.5 ^M. A separate LinkedIn post claims Microsoft's MAI-Thinking-1, after tuning for McKinsey, outperformed OpenAI's GPT-5.5 at a 10x better cost efficiency, though this claim comes from a user-generated source and should be treated with caution rather than as a confirmed independent benchmark ^L.

The Broader Strategy: Multi-Model, Not Replacement

The model swap in Excel and Outlook does not mean Microsoft is abandoning OpenAI or Anthropic. The company continues to sell both partners' models through Azure and has itself integrated Anthropic's Claude models into some Office 365 Copilot features as recently as May 2026 ^GPA. What is changing is Microsoft's model mix: instead of relying solely on OpenAI for Copilot workloads, the company is moving toward a flexible, multi-model strategy where in-house MAI models handle high-volume, cost-sensitive tasks in Excel and Outlook, Anthropic Claude handles certain creative or analytical tasks in PowerPoint and Word, and OpenAI models remain available for other use cases ^DTT.

According to sources cited by Bloomberg, Microsoft's internal goal is to reduce what it pays third-party AI providers by shifting as much inference demand as possible to its own models — a strategy that becomes easier to execute now that Microsoft controls both the infrastructure (Azure) and the model layer (MAI) ^BW.

What It Means for Microsoft's $13 Billion OpenAI and $5 Billion Anthropic Investments

Microsoft has invested roughly $13 billion in OpenAI and $5 billion in Anthropic, according to multiple reports ^KCSG. Those investments remain strategic: Microsoft still resells both partners' models via Azure, and OpenAI alone has committed to approximately $250 billion in Azure consumption ^Y. However, the MAI push gives Microsoft cost leverage over those partners. Every Copilot prompt that moves from an external model to an MAI model is a prompt for which Microsoft keeps the margin instead of paying a third-party API fee ^DWK.

Microsoft AI CEO Mustafa Suleyman told Bloomberg that "our goal is to reduce and ultimately eliminate" payments to Anthropic ^I. The same principle applies to OpenAI: Microsoft's own models are now an economic option, reducing the urgency of accepting price increases from either partner ^WI.

The IPO Question

The provided sources confirm that both OpenAI and Anthropic filed confidentially for IPOs in early June 2026 ^CL. However, the evidence does not support a firm conclusion about how Microsoft's MAI shift will affect those IPOs. What can be said is narrower: Microsoft's use of MAI in Excel and Outlook may reduce some inference demand that would otherwise have generated revenue for OpenAI or Anthropic ^DB. If that pattern expands across other Microsoft products, it could weaken the assumption that Microsoft Office and Copilot usage will automatically translate into growing enterprise inference revenue for outside model providers ^DTM. The sources do not quantify the revenue impact on either company, nor do they establish how public-market investors would price that risk.

Key Takeaway

Microsoft's MAI push is a cost-reduction and dependency-reduction strategy, not a full divorce from its AI partners. The confirmed use of MAI in Excel and Outlook, the launch of seven MAI models at Build 2026, and the reported pricing advantages versus comparable third-party models all point to a future where Microsoft owns more of its AI stack — while still keeping one foot in the OpenAI and Anthropic ecosystems ^DMBMB.