Microsoft is Developing its Own Models: MAI's Seven Launches, a Sign of Breaking Up with OpenAI?

If you've been in the tech industry, you know that some product launches are actually about "shifting power structures" beneath the surface. The Microsoft Build 2026 developer conference on June 2 was one such event.

Microsoft unveiled seven self-developed MAI series AI models at once. The focus isn't on the quantity, but on the repeated emphasis that these models are "completely independent of OpenAI technology, trained from scratch." This is noteworthy, given that Microsoft has invested billions of dollars in OpenAI and integrated GPT into Copilot, Office, and Azure. The fact that Microsoft is now developing its own models is more intriguing than the model specifications themselves.

Event Background

Microsoft's relationship with OpenAI has been one of the most delicate "frenemies" in the tech circle in recent years. As OpenAI's largest investor and cloud provider, Microsoft's products heavily rely on GPT. However, no tech giant wants to put its core capabilities solely on "someone else's model" – what if the prices rise, the supply is cut off, or the partner decides to compete directly?

Therefore, Microsoft established its own AI department (MAI, Microsoft AI) and began quietly developing its models. The seven launches at Build 2026 are the first large-scale showcase of this effort.

Key Highlights

According to Microsoft's official announcements and media reports, the key highlights of this launch include:

• MAI-Thinking-1: Microsoft's first self-developed reasoning model, featuring a sparse MoE architecture, 350 billion active parameters, and a 256K context window. It emphasizes "training from scratch, using commercial licensed data, without distilling any third-party models." In the AIME 2025 and AIME 2026 math competitions, it scored 97.0% and 94.5%, respectively. In a blind test by independent evaluation partner Surge, its performance was rated superior to Claude Sonnet 4.6.
• MAI-Code-1-Flash: A lightweight coding model with 50 billion parameters, which started rolling out on GitHub Copilot (VS Code) on June 2, covering Copilot's Free, Pro, Pro+, and Max plans. Microsoft claims it outperforms Claude Haiku 4.5 in four core programming tests, leading by 16 percentage points (51.2% vs. 35.2%) in SWE-Bench Pro.
• Other models: The seven launches also include MAI-Image-2.5 for image generation, MAI-Transcribe-1.5 for speech-to-text supporting 43 languages, and the MAI-Voice-2 speech model.

Microsoft positions this suite as a step to "reduce developer costs and decrease dependence on a single supplier."

Market Impact Analysis

For Taiwanese users: In the short term, you may not directly feel the impact of MAI models, as they will be operating behind the scenes in Copilot, Office, and Azure. However, in the long run, this is a positive development – when Microsoft has its own models to compare and substitute with OpenAI, the cost pressure on the entire ecosystem will decrease, potentially reflecting in subscription prices or free quotas. For Taiwanese engineers and office workers who use Copilot, having an additional model option and reducing the risk of being tied to a single model is a good thing.

For enterprise applications: This serves as a reminder to CIOs and technical directors that "not putting all eggs in one basket" applies not only to investments but also to AI suppliers. Since Microsoft is developing its own models for redundancy, companies introducing AI should consider the architecture of "model replaceability" rather than binding their systems to a specific API. This is why tools like OpenCode that do not bind models are gaining popularity.

For developers: The most noticeable impact is MAI-Code-1-Flash being directly integrated into GitHub Copilot. Having an additional fast and low-cost coding model option is beneficial for daily completion and small tasks. However, it's essential to note that officially published benchmark numbers should be taken with a grain of salt – preferences in blind tests and leads in specific test sets may not necessarily align with the actual feel in your projects. The true performance can only be determined by hands-on experience.

Future Development Trends

The real signal from this event is that the AI competition landscape is shifting from "whose model is the strongest" to "who can use models most efficiently and integrate them deepest." As the capability gap between top models narrows, giants will compete on costs, ecosystem binding, and the ability to seamlessly integrate AI into daily software. Microsoft's self-developed models are a move to grasp its own cards and avoid being strangled in this "integration war."

It can be expected that Google, Amazon, and even Apple will find a balance between developing their own models and purchasing necessary ones. For users, this multi-strong competition is the best scenario – no single company can dominate, and prices and innovations will be driven forward.

TheAI Academy Summary and Commentary

Microsoft developing its own models isn't because OpenAI's models are not useful, but because "putting lifelines in someone else's hands" makes any mature giant uneasy. MAI's seven launches are more of a clear strategic statement: "I have a backup plan now."

Commentary: Don't focus too much on benchmark numbers that show who wins by a few points; those are marketing figures. What's truly important is that AI is transitioning from "a solo show of star models" to "an integration war of giant ecosystems." For users like us, multi-strong competition means cheaper options and more choices, which is a good thing.

Practical advice for Taiwanese readers: If you're a developer, try the new MAI-Code-1-Flash in GitHub Copilot and compare its actual performance with Claude and Cursor; don't just trust official numbers. If you're a corporate decision-maker, consider "model replaceability" as a basic principle for introducing AI. To understand the full picture of this model turmoil, you can read about the 2026 model wave and AI Agent trends.

Data Sources

• Microsoft AI official blog: Introducing MAI-Code-1-Flash
• Microsoft AI: Launching seven new MAI models
• CNBC, TechTimes, Neowin, and other media reports

This article is compiled based on publicly available information as of June 2026, with model specifications, performance data, and launch ranges following Microsoft's official announcements.