Microsoft kicked off Build 2026 at Fort Mason Center in San Francisco on June 2, unveiling a family of seven in-house AI models under its MAI brand — the largest single-day expansion of Microsoft's proprietary model catalog — with a flagship reasoning model explicitly built to be independent of the company's $13 billion partner, OpenAI.

MAI-Thinking-1: No OpenAI DNA

The headlining release is MAI-Thinking-1, which Microsoft describes, in the words of its official blog post by GitHub and Developer Marketing chief Kyle Daigle, as "trained from scratch with zero distillation on enterprise grade, clean and commercially licensed data you can build on with confidence." The model is a sparse Mixture of Experts architecture with 35 billion active parameters and a 256,000-token context window — large enough to process a 600-page document in a single inference pass.

The decision to train without OpenAI data or model distillation is notable because it directly addresses IP provenance concerns that have complicated enterprise adoption of models derived from third-party sources. According to Microsoft's primary announcement page for the model, MAI-Thinking-1 scores 97.0 percent on AIME 2025 and 94.5 percent on AIME 2026, mathematical reasoning benchmarks that have become a standard frontier-model yardstick. On SWE-Bench Pro — a software engineering evaluation — the company says it matches Claude Opus 4.6 on coding tasks, and that independent raters hired through Surge (Microsoft's human evaluation partner) preferred it over Claude Sonnet 4.6 in blind side-by-side tests. Independent replication of those benchmark figures has not yet occurred, and the claims should be treated as self-reported until confirmed by external labs.

MAI-Thinking-1 entered private preview on Microsoft Foundry on June 2. It supports function calling, multi-step instruction following, and is compatible with the standard Chat Completions API.

The April Partnership Renegotiation Provides Context

The timing is deliberate. In April 2026, Microsoft and OpenAI amended their partnership agreement, according to reporting covered by The Register and CNBC. The renegotiation ended Microsoft's exclusive license to OpenAI intellectual property and removed Microsoft's revenue-share obligation to OpenAI, while preserving OpenAI's capped revenue share to Microsoft through 2030. Microsoft CEO Satya Nadella summarized the new stance at the Build keynote: "We believe the time has come for every company to just move from consuming a frontier model to fully participating at the frontier in the frontier ecosystem."

Mustafa Suleiman, CEO of Microsoft AI, went further, claiming that after fine-tuning Microsoft's own models for McKinsey's specific workflows, the company achieved 10 times better cost efficiency compared to OpenAI's GPT-5.5 on the same task, according to CNBC's reporting from the keynote. That claim has not been independently reproduced.

The practical upshot: Microsoft's model strategy has shifted from distributing OpenAI's outputs to building parallel in-house capability at every major tier of the AI developer stack.

Six More MAI Models Round Out the Family

Alongside MAI-Thinking-1, Microsoft announced:

  • MAI-Code-1-Flash — An inference-efficient coding model now rolling out across all GitHub Copilot tiers (Free, Pro, Pro+, and Max) and available in Visual Studio Code. The model was trained inside GitHub Copilot's production harness rather than benchmarked externally, which Microsoft says improves reliability in agentic coding workflows. Some post-keynote reporting described the model under the name "Project Polaris" and suggested it would replace GPT-4 Turbo as the default Copilot engine in August 2026, but Microsoft's official Build materials do not confirm that timeline.

  • MAI-Image-2.5 and MAI-Image-2.5 Flash — Microsoft's first models handling both text-to-image and image-to-image workloads, ranked third and second respectively on the Arena AI leaderboard. Already live in PowerPoint and rolling out to OneDrive.

  • MAI-Transcribe-1.5 — Speech recognition supporting 43 languages, with a claimed five-times speed improvement over competing transcription models and streaming support coming soon.

  • MAI-Voice-2 — Speech generation expanded to more than 15 additional languages with new voice options.

All MAI models are being distributed beyond Microsoft's own platform: Fireworks AI (now generally available on Microsoft Foundry), Baseten, and OpenRouter are listed as third-party distribution channels.

OpenAI Simultaneously Goes Live on AWS — Microsoft's Former Exclusive

A parallel development underscores the new competitive landscape. As of June 2, OpenAI's GPT-5.5 and GPT-5.4 models, along with its Codex coding agent, are generally available on Amazon Bedrock — Microsoft's rival cloud. According to the AWS announcement page, pricing matches OpenAI's direct rates with no additional fees, and usage counts toward customers' existing AWS commitments. The April 2026 partnership amendment that ended Microsoft's exclusive OpenAI license is what made this possible; OpenAI is now free to distribute through Amazon's infrastructure, which more than 100,000 organizations already use.

This creates a structurally different AI market from the one that existed six months ago: Microsoft is now competing against its own investee's models on a rival cloud, while also building proprietary alternatives to those models.

Infrastructure Note: Claude in Foundry Still Runs on Anthropic Servers

For enterprise teams evaluating Microsoft Foundry as a model-neutral orchestration layer, one detail from the Build announcement deserves attention. Claude models — including the newly released Claude Opus 4.8 (launched May 28) — remain available in Foundry under unified Azure billing and Microsoft Azure Consumption Commitment (MACC) eligibility, according to Anthropic's documentation. However, they currently run on Anthropic-managed infrastructure, not on Azure's regional compute. That is a different operational profile from OpenAI or MAI models, which run natively on Azure. Anthropic lists EU-native inference support in Foundry as "Coming 2026" with no specific date, meaning organizations with strict EU data residency requirements may need supplementary compliance planning if they deploy Claude through Foundry.

Majorana 2: Quantum Claims Under Scrutiny

Microsoft closed the Build keynote with Majorana 2, a next-generation quantum chip claiming an average qubit lifetime of 20 seconds — 1,000 times higher reliability than its previous generation — and a target of one million qubits on a single chip by 2029. Those claims rest on a preprint that has not been peer-reviewed. Scientific American reported that independent physicists including Henry Legg of the University of St. Andrews said the underlying results "would never make it through peer review" by an independent group, citing that the paper presents only Z measurements rather than the full X and Z measurements needed to demonstrate a functioning topological qubit. Microsoft's history with this technology adds caution: a 2018 claim about Majorana zero modes was formally retracted.

The Competitive Picture

Build 2026's underlying message is that Microsoft is no longer solely a distributor of other companies' models. With MAI-Thinking-1, it has a reasoning model at a competitive benchmark tier; with MAI-Code-1-Flash, it has a coding model integrated into 150 million GitHub Copilot users' default tooling; with MAI-Image-2.5, it has generation models already deployed in PowerPoint and OneDrive. The company is positioning Microsoft Foundry and GitHub Copilot as the orchestration layer above all of this — a multi-model platform where enterprises route workloads across MAI, OpenAI, and Anthropic models under unified Azure billing and governance, regardless of who built the underlying intelligence.