Microsoft Launches Seven In-House MAI Models, Built From Scratch Without Distillation to Cut OpenAI Reliance

Overview

Microsoft announced seven of its own AI models on June 2, 2026, the company’s most concerted push yet to build a frontier-model stack in house rather than rely on OpenAI. According to GIGAZINE, “On June 2, 2026, Microsoft announced seven of its proprietary AI models.” By the time CEO Satya Nadella took the stage at Build in San Francisco that day, the company had already rolled out the seven in-house models, as reported by Technobezz.

The release is the work of Microsoft AI, the consumer-facing division run by Mustafa Suleyman, who framed the launch in a Microsoft AI blog post as a step toward “long term self-sufficiency for Microsoft and our partners.”

What We Know

The seven models span reasoning, coding, image generation, transcription, and voice. According to Microsoft AI, the lineup is MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2, MAI-Voice-2-Flash, and MAI-Image-2.5-Flash. GIGAZINE independently lists the same set of seven names.

The flagship is MAI-Thinking-1, a reasoning model that the MAI-Thinking-1 model page describes as a “35B-active, ~1T-total parameters, sparse Mixture of Experts model” that “supports long context with a 256k token window.” The same page reports the model “reaches 97.0% on AIME 2025, and 94.5% on AIME 2026” and runs “toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro,” while “users preferred MAI-Thinking-1 over Claude Sonnet 4.6” in evaluations. Technobezz corroborates the headline specifications, describing MAI-Thinking-1 as a model “with 35 billion active parameters” and “a 256,000-token context window.”

The central message of the launch is that the models were built without leaning on rivals’ systems. Microsoft AI states, “We train our reasoning models from scratch. We don’t distill from other labs and we don’t rely on unlicensed or opaque data,” according to the Microsoft AI blog post. The MAI-Thinking-1 page reiterates that the model was “trained without distillation from third party models” and “trained it from the ground up on clean, traceable and enterprise-grade data.” Technobezz adds that all the models were “trained from scratch on Azure infrastructure using commercially licensed data, with no distillation from third-party systems.”

Microsoft is wiring the models directly into its developer surfaces. MAI-Code-1-Flash, a coding model with “5 billion active parameters,” is “tailor-made for and deeply integrated into GitHub Copilot, VS Code and the Microsoft stack,” according to Microsoft AI. The company also points to its own silicon: the same post says Microsoft co-designs with its “Maia 200 silicon, and are already seeing a 1.4x efficiency boost.” MAI-Thinking-1 is “available in private preview on Microsoft Foundry today,” the model page says.

The self-sufficiency framing carries a direct competitive subtext. IndexBox reports that Microsoft’s models “surpassed OpenAI’s GPT 5-5 while achieving a tenfold reduction in costs” against McKinsey benchmarks, a claim echoed by Technobezz, which says “the company outperformed OpenAI’s GPT-5.5 on quality with what it projects as ten times better cost efficiency.” According to IndexBox, Microsoft has “adjusted its agreement with OpenAI, placing a ceiling on revenue-sharing payments and terminating its exclusive right to market OpenAI’s models.” Both Microsoft AI and IndexBox note the models will also be reachable through third-party platforms, naming OpenRouter, Fireworks, and Baseten.

What We Don’t Know

Microsoft has published headline benchmark figures, but independent, third-party evaluations of MAI-Thinking-1 against Claude and GPT-5.5 had not been released at announcement. The reported “tenfold reduction in costs” is drawn from Microsoft’s own framing against McKinsey benchmarks rather than a neutral arbiter, and the per-token pricing for the new models was not detailed in the launch materials reviewed here. The full terms of the revised OpenAI agreement — including how the revenue-sharing ceiling is calculated — were not disclosed.

Analysis

The strategic stake is larger than any single benchmark. Technobezz notes that Microsoft “invested $13 billion in OpenAI across multiple tranches beginning in 2023, securing exclusive cloud rights on Azure.” Standing up a from-scratch model family that Microsoft says rivals Claude Opus 4.6 on coding gives the company an internal alternative for the workloads it currently routes to OpenAI. Suleyman situates the effort in a longer arc, writing in the Microsoft AI post that the goal is “to build what we think of as a hill-climbing machine: an organization that can continuously improve, cycle after cycle,” with an “ultimate goal” he calls “Humanist Superintelligence.”