Microsoft Build 2026: MAI-Thinking, Copilot Super App and the first proprietary models

Microsoft Build 2026: MAI-Thinking, Copilot Super App and the first proprietary models

🔎 Why Microsoft is finally building its own AI models

After four years and $13 billion injected into OpenAI, Microsoft drew a red line at Build 2026. The company unveils MAI-Thinking-1, its first proprietary reasoning model, alongside a full range of in-house models. This is not just a simple product update. It is a geopolitical signal in the AI industry.

On April 2, 2026, Microsoft had already released three foundational models under the MAI brand. The signal was clear: the dependence on OpenAI had an expiration date. Build is the confirmation that this strategy was not a test but a large-scale execution plan.

Selon The Verge, this edition of Build marks a turning point: Microsoft is no longer content to be OpenAI's premium distributor. It is becoming a full-fledged modeling player, with a vertical stack ranging from Maia silicon to models, all the way to the Copilot interface.

The essentials

• MAI-Thinking-1: Microsoft's first proprietary reasoning model, built from scratch by Mustafa Suleyman's team, distinct from the GPT lineage.
• Expanded MAI range: MAI-Image-2.5, MAI-Transcribe-1.5 and multilingual MAI-Voice-2 complete the multimodal stack.
• Copilot Super App ("One Copilot"): a single interface unifying AI chat, coding (GitHub Copilot) and browsing, led by Jacob Andreou, with a launch planned for late summer 2026.
• Project Polaris: in-house model that will replace GPT-4 Turbo as GitHub Copilot's default engine starting in August 2026, with 100K lines of context and inference on Maia accelerator.
• Independence strategy: the gradual break with OpenAI is official. Microsoft is building its own vertical AI stack.

Recommended tools

MAI-Thinking-1 : Microsoft's first true in-house reasoning model

MAI-Thinking-1 is not a fine-tune of GPT-4. It is a reasoning model built from scratch by Mustafa Suleyman's team, recruited in 2024 to lead Microsoft's consumer AI division. The distinction is fundamental.

According to Digit.in's analysis, MAI-Thinking-1 positions itself as a "reasoning AI" model — it is not designed to generate fluent text, but to break down complex problems step by step. It shares the same philosophy as OpenAI's "o1" models, but with its own architecture.

The model is already partially accessible. MAI-DS-R1 is available on OpenRouter in a free version, with improved performance on harm mitigation benchmarks and general reasoning tasks. This is a strong signal: Microsoft wants developers to test its stack without friction.

For Azure developers, this means a concrete alternative to OpenAI models for reasoning-heavy use cases: complex log analysis, multi-file debugging, architectural planning. If you use free AI models without sacrificing quality, MAI-DS-R1 on OpenRouter is a serious option.

Is it a GPT killer?

Not yet. The June 2025 general LLM benchmark places Gemini 3.1 Pro (92) and GPT-5.5 (91) at the top. MAI-Thinking-1 does not appear on it yet, suggesting either that its performance is still below the top tier, or that it has a different positioning (pure reasoning vs. general generation).

On the other hand, in the agentic benchmark, GPT-5.5 dominates at 98.2, followed by Gemini 3 Pro Deep Think (95.4) and Claude Opus 4.7 Adaptive (94.3). If MAI-Thinking-1 is targeting this segment, it must prove it can rival these scores. For now, it is a promising model but not disruptive on paper.

The real threat to OpenAI is not MAI-Thinking-1's score. It's the ecosystem. If Microsoft pushes its models through Azure, GitHub Copilot, and Windows, the volume of usage will do the rest — even with a slightly inferior model.

The complete MAI lineup: image, voice, and transcription

Microsoft did not unveil a single model. The MAI lineup presented at Build 2026 covers four distinct modalities, indicating a full-stack strategy rather than a single-product approach.

TestingCatalog reports that Microsoft was specifically preparing MAI-Voice-2 and MAI-Image-2.5 for this conference, with significant improvements over previous generations.

MAI-Image-2.5 replaces dependencies on DALL-E for image generation within the Microsoft ecosystem. Direct integration into Copilot and Windows design tools eliminates a costly API call to OpenAI.

MAI-Transcribe-1.5 targets speech-to-text processing, a segment where OpenAI's Whisper has dominated so far in the Microsoft stack. According to announcements, version 1.5 improves accuracy on technical jargon and accents.

MAI-Voice-2 is the most strategic model after MAI-Thinking-1. Its multilingual capability positions Microsoft directly against the voice offerings from OpenAI and Anthropic. Winaero confirms that this model was one of the pillars of the Build 2026 keynote.

These models are not gimmicks. Each and every one of them replaces an OpenAI API call in the Microsoft stack. This is systematic replacement, not showcase innovation.

One Copilot : the super-app that unifies everything

The most ambitious project at Build 2026 is not a model. It's an interface. Fortune revealed that Jacob Andreou, former vice president at Stripe, is leading the development of a Copilot "super app" that unifies chat, coding, and browsing.

Alexis Tauzin analyzes this initiative under the name "One Copilot": a single application that merges GitHub Copilot, AI chat, Copilot Cowork, and a new agentic layer called "Autopilot". The launch is scheduled for late summer 2026.

The stakes are considerable. Today, Microsoft's AI experience is fragmented: a Copilot tab in Edge, an extension in VS Code, a chatbot in Windows, agents in Teams. One Copilot aims to eliminate this fragmentation.

Up against ChatGPT and Claude

ChatGPT already has a super-app with its unified interface (chat, code interpreter, browsing, canvas, plugins). Claude offers a streamlined but single-window experience. One Copilot wants to go further by natively integrating coding — not just a simple code editor within a chat, but the full GitHub Copilot inside the interface.

The strategic difference is clear: ChatGPT and Claude are SaaS products. One Copilot is an ecosystem product. It leverages Windows, Azure, GitHub, Office — assets that no competitor possesses.

Autopilot : the agentic layer

The most interesting detail is the "Autopilot" layer mentioned in Tauzin's analysis. It is no longer about assistance (copilot) but autonomy (autopilot) — the AI executes multi-step tasks without constant supervision. This is exactly the segment where the agentic benchmarks of June 2025 show GPT-5.5 (98.2) and Gemini 3 Pro Deep Think (95.4) at the top.

If Autopilot relies on MAI-Thinking-1 on the backend, Microsoft controls the entire chain: model, agent, interface. This is the vertical stack that only Google can also claim to assemble.

Project Polaris : GitHub Copilot's new engine

AI Tool Briefing révèle that Microsoft Project Polaris will replace GPT-4 Turbo as GitHub Copilot's default model in August 2026. The specs are impressive: 100,000 lines of context and optimized inference on Microsoft's Maia accelerators.

This is the most concrete signal of the transition. GitHub Copilot, Microsoft's most profitable AI product, will migrate to an in-house model. Not an OpenAI fine-tune — a model built and optimized internally.

The advantage of the 100K lines of context is major for development. Current benchmarks place DeepSeek V4 Pro (Max) at 88 in general and GPT-5.3 Codex at 87, but neither natively offers this context window optimized for code within the Microsoft ecosystem.

For developers who configure models and providers in Hermes Agent, the arrival of Polaris via the Azure API means a new provider to integrate — with potentially lower costs thanks to Maia silicon optimization.

If you already use the best AI tools for code like Cursor or Copilot, the transition to Polaris should be seamless. But it will mark the moment when your coding assistant is no longer powered by OpenAI.

The independence strategy: why now?

Tech-Insider documents the historical context: on April 2, 2026, Microsoft released three foundational proprietary models under the MAI brand. This was, according to their analysis, "the clearest signal of a break with the $13 billion OpenAI partnership."

Several factors explain this timing:

Financial pressure. Every API call to OpenAI costs Microsoft dearly. With Azure margins under pressure and Anthropic reaching a $47 billion revenue run-rate and surpassing OpenAI in certain segments, Microsoft can no longer afford to be just a reseller.

The risk of dependency. OpenAI is negotiating with other cloud providers. The exclusive relationship has been dead since 2024. Microsoft must be able to offer a complete AI stack without depending on a partner that could become a direct competitor.

Maia silicon. Microsoft's custom AI accelerators are ready. Without in-house models, these chips have no optimized workload. MAI-Thinking-1 and Project Polaris are the first products designed natively for Maia.

Google's competition. Gemini 3.1 Pro dominates general benchmarks at 92 points. Google owns its own vertical stack (TPU, Gemini, Android, Chrome). Microsoft must match this integration or risk becoming the "Windows of AI" facing the "Google integrated experience".

Calipia reports that according to The Information and Reuters, Microsoft has been preparing this counter-attack for months, with a range of in-house AI models including a specific coding model for GitHub Copilot — that is Project Polaris.

Concrete impact for Azure developers

If you deploy on Azure today, your AI stack probably looks like this: Azure OpenAI Service for LLMs, Whisper for transcription, DALL-E for images, OpenAI embeddings for vector search. Each of these calls is billed by OpenAI, with a margin taken by Microsoft.

With the MAI lineup, the math changes:

• MAI-Thinking-1 replaces "o1" calls for complex reasoning
• MAI-Voice-2 replaces Whisper
• MAI-Image-2.5 replaces DALL-E
• Project Polaris replaces GPT-4 Turbo in code workflows

Prices are not yet published for all models, but the economic logic is undeniable: a model optimized for Maia silicon is cheaper to serve than an OpenAI model running on rented NVIDIA GPUs. Microsoft can either lower prices or improve its margins.

For teams that use free models without sacrificing quality, the availability of MAI-DS-R1 on OpenRouter in a free version is a preview of this strategy: make MAI models accessible to build the habit, then monetize via Azure for production.

Migration in practice

The migration won't be instant. OpenAI remains Azure's default provider for many use cases. But for everything related to reasoning, voice, and images in the Microsoft ecosystem, the switch to MAI is planned.

Developers need to start abstracting their AI calls behind a provider-agnostic layer. This is exactly what Hermes Agent allows: configure multiple providers and switch between OpenAI and MAI without rewriting the application code.

Vaporware or a genuine pivot? The verdict

Abit.ee asks the right question in its analysis: distinguishing the real from the vaporware in the Build 2026 announcements. The answer is nuanced.

What is real: MAI-DS-R1 is already on OpenRouter. The foundational MAI models were launched on April 2, 2026. Project Polaris has a migration date (August 2026) for GitHub Copilot. These are not concepts — these are products with timelines.

What looks like vaporware: One Copilot / Autopilot has no precise public availability date ("late summer 2026" is vague). The unified super-app is a complex project that involves merging existing interfaces with different teams and codebases. Jacob Andreou is talented, but the organizational challenge is immense.

The verdict: Microsoft is executing a real but gradual pivot. The model stack (MAI) is concrete. The interface stack (One Copilot) is ambitious and risky. The success of the former makes the latter possible, but does not guarantee it.

❌ Common mistakes

Mistake 1: Confusing MAI-Thinking-1 with a GPT fine-tune

MAI-Thinking-1 is built from scratch. It is not GPT-4 with a special system prompt or custom RLHF. The architecture is proprietary. Making this confusion means missing the entire strategic significance of the announcement.

Mistake 2: Thinking Microsoft will burn bridges with OpenAI

The break is gradual, not abrupt. Azure OpenAI Service remains a flagship product. GPT models will remain available. Microsoft is diversifying its stack, not replacing it overnight. The two will coexist for at least 18-24 months.

Mistake 3: Ignoring Maia silicon in the analysis

Without Maia accelerators, MAI models have no competitive advantage. It is the combination of model + silicon + Azure infrastructure that creates differentiation. Analyzing MAI-Thinking-1 without mentioning Maia is analyzing half the story.

Mistake 4: Underestimating the super-app challenge

Unifying chat, coding, and browsing into a coherent interface is a daunting UX problem. ChatGPT took two years to get there. Microsoft starts with more products to merge. The risk of a "Frankenstein interface" is real.

❓ Frequently Asked Questions

Does MAI-Thinking-1 replace GPT-5.5 on Azure?

No. MAI-Thinking-1 positions itself as an alternative for reasoning use cases, not as a generic replacement for GPT-5.5. Both will coexist on Azure with differentiated use cases.

When will One Copilot be available?

Microsoft is targeting "late summer 2026", but no exact date was announced during Build. Expect a limited beta first, followed by a gradual rollout via Windows Update.

Will Project Polaris be better than GPT-5.3 Codex for code?

On paper, Polaris's advantage is its 100K-line context window optimized for code, whereas GPT-5.3 Codex (87 on the general benchmark) does not offer this contextual specialization. But real-world coding benchmarks are needed to confirm.

Can I test the MAI models now?

Yes. MAI-DS-R1 is available for free on OpenRouter. For the other models (full MAI-Thinking-1, MAI-Image-2.5, MAI-Voice-2), access is through Azure with a likely waitlist.

What impact on Azure OpenAI pricing?

In the short term, no changes. In the medium term (12-18 months), expect downward pressure on OpenAI prices as Microsoft pushes its MAI models as a cheaper alternative.

✅ Conclusion

Microsoft Build 2026 marks the end of the "Microsoft = OpenAI distributor" era. With MAI-Thinking-1, Project Polaris and One Copilot, the company is assembling a vertical AI stack that has only one equivalent: Google with Gemini + TPU + Android. The models are not yet at the level of GPT-5.5 or Claude Opus 4.7, but the ecosystem strategy is in place — and that's what matters. For developers, the message is clear: start abstracting your AI calls, because Azure's default provider in 18 months may no longer be OpenAI.