Microsoft Breaks Free from OpenAI and Unveils Its AI Ambitions at Build

At Microsoft's annual Build conference, the company unveiled a series of new initiatives in artificial intelligence, signaling a significant break from OpenAI. Among the announcements were a super app, internal reasoning models, a cybersecurity tool, and AI agents similar to OpenClaw. These developments mark Microsoft's intent to establish itself as a major player in the AI field.

For years, Microsoft has heavily relied on its exclusive partnership with OpenAI for its AI activities. However, this collaboration has evolved into an effective separation at the end of April, although Microsoft remains OpenAI's primary cloud partner for now. This year's Build conference was seen as a declaration of independence, with CEO Satya Nadella stating that these events represent an opportunity to "become aware of the new opportunity."

Mustafa Suleyman, head of AI at Microsoft, was even more explicit during an interview. "The goal is to prove that we can become one of the top four labs in the world," he said. He emphasized that Google DeepMind, OpenAI, and Anthropic are currently the leaders, but Microsoft aspires to join them. Suleyman expressed his intention to build the best cutting-edge models in the world, fully multimodal, and to demonstrate that Microsoft can innovate without simply borrowing ideas from others.

MAI-Thinking-1: An Ambitious Reasoning Model

One of Microsoft's first initiatives at Build was to close its gap in the AI model space. Suleyman introduced MAI-Thinking-1, the company's first reasoning model, along with six other new models focused on image, voice, transcription, and programming. The MAI-Thinking-1 model, of medium size, is primarily aimed at enterprise customers and was "built from the ground up for serious deployments in mathematics, programming, and the real world." Microsoft is several years behind OpenAI and Anthropic, with OpenAI having started releasing reasoning models as early as fall 2024. However, Suleyman highlighted MAI-Thinking-1's performance on benchmarks such as programming, as well as its lower cost compared to OpenAI's equivalents for certain tasks, which is crucial at a time when companies are looking to cut costs.

Although Microsoft has had years to learn from OpenAI, Suleyman clarified that the development of MAI-Thinking-1 did not involve distillation, meaning it was not trained using another company's AI model. If MAI-Thinking-1 proves to be effective, Microsoft clearly wants this to be attributed to its own efforts and not to OpenAI's influence.

Suleyman stated that for Microsoft, "the turning point was the renegotiation of our contract with OpenAI. This meant we were allowed to train models at a larger scale and to explicitly pursue superintelligence entirely with our own intellectual property, with our own data, without distillation, training from scratch."

Cybersecurity and AI Agents: Microsoft Positions Itself

Nadella also highlighted the recently launched AI cybersecurity tool, MDASH, which brings together 100 AI agents to detect exploitable bugs "better than any single model." This statement clearly targeted Claude Mythos Preview, introduced by Anthropic in April, whose access was expanded just before Build. OpenAI also has its own cybersecurity-focused system, and all three companies are positioning themselves to capture government and enterprise markets.

Microsoft finds itself in a complex situation with AI agents. The popular open-source platform OpenClaw has shown a potential path for AI agents, and after OpenAI quickly hired its creator, Peter Steinberger, Microsoft (along with other companies) is looking to catch up. One of Microsoft's key strategies is to ensure that OpenClaw works well with Windows. At Build, Nadella reaffirmed his commitment to supporting OpenClaw, and Microsoft employees engaged with developers present on how they were using it.

Steinberger himself made a surprise appearance, eliciting a strong reaction from the audience, and touted how OpenClaw had enhanced its security and gained user trust. "What I kept hearing was: 'Peter, I love my Claw, but can I use it at work?'" Steinberger said. "You can totally run OpenClaw within your company now, and we've even made the harness itself a plug-in." Steinberger added that it doesn't matter if someone trusts Copilot, Codex, or another company's programming platform; users can now run OpenClaw on top of that via Windows.

Copilot: Microsoft's Super App

Microsoft is also promoting its own super app, Copilot, which integrates agents similar to OpenClaw. A super app is a major focal point for OpenAI right now — President Greg Brockman is leading the development of an app that will link ChatGPT, the Codex programming platform, and the Atlas web browser. Microsoft's strategy is similar, gathering a variety of existing Copilot AI assistants. Its agents, called Autopilots, are designed to act as a useful user interface. Cassidy Williams, Senior Director of Developer Engagement at GitHub, described Copilot as "the development and operations base on your computer," demonstrating how multiple agents could perform tasks like application creation.

The Autopilots are specifically designed to appeal to enterprise customers — Nadella referred to them as "autonomous, long-lasting agents with full compliance to enterprise requirements." The first one Microsoft will offer is Scout, described as "your always-active personal agent," but customers can build and customize their own. Autopilot agents are expected to be able to browse an inbox, join group discussions in Teams, check a calendar, and send daily briefings, among other tasks. Consequently, employees on stage at Build repeatedly emphasized Copilot's security tools and safeguards — clearly aiming to reassure enterprise customers who may have heard horror stories about tools like OpenClaw.

Suleyman made sure to emphasize, time and again, Microsoft's "humanistic superintelligence" as an "AI that prioritizes humanity first" — part of the recent rebranding of AI companies to make AGI less frightening at a time when people are increasingly wary of AI.

Nvidia CEO Jensen Huang, another speaker known for his close collaboration with OpenAI, appeared via video call to tout how Nvidia's RTX Spark chip powers Microsoft's AI agent goals. "I could be traveling, I'm on the phone, and I can send a message to my PC... and that would activate the tools on the PC," Huang said. "The idea that the PC has evolved from a personal computer to a personal AI is really exciting."

Microsoft has spent years betting on OpenAI, and in some ways, this has left it behind in the AI race. But as OpenAI and other competitors turn to the enterprise to finally generate revenue, Microsoft has clear advantages. The company already has a substantial customer base and, compared to other AI companies, a reputation for security. And like Google, it also has considerable financial resources, significant computing resources, and a diversified revenue stream, allowing it to make big bets without too much risk.

Suleyman told The Verge: "There are a lot of people chasing startup valuations or about to go public, so we can operate with a bit more humility and long-term optimization." He added: "We have the money to buy Anthropic models when we need them. We have optionality in Azure with 11,000 models, so people can literally use whatever they want whenever they want, but that gives us the time to do things right from the start."

At the same time, many questions remain unanswered. Microsoft has highlighted numerous successes and advancements for its seven new models, but this does not always translate into real-world adoption, and even a new model that takes the lead for a week or two can quickly fall behind. AI super apps are still a largely unexplored idea. And Microsoft is entering a crowded but still largely disappointing AI agent market with a product we have yet to see in action. There is still plenty of room for its promises to fail.