Microsoft Just Launched 7 AI Tools at Once. Here Is What Each One Means for You.

MAI-Thinking-1, MAI-Voice-2, MAI-Image-2.5, a chip that beats Nvidia on cost, and an always-on agent. Every claim verified from official Microsoft Build 2026 sources.

• Microsoft held its annual Build developer conference on June 2 and 3, 2026 in San Francisco. In a single keynote, CEO Satya Nadella and AI chief Mustafa Suleyman launched seven new in-house AI models, an always-on productivity agent, a custom chip that undercuts Nvidia on price, and a platform for an entirely new class of AI-first devices.
• None of these models use OpenAI technology. Microsoft trained every one of them from scratch, on its own commercially licensed data, without distillation from any third-party model. That is a significant shift for a company that built its AI business on top of OpenAI's GPT models.
• Here is every announcement, what the verified numbers actually say, and what it means if you are already paying for similar tools.

MAI-Thinking-1: Microsoft's First In-House Reasoning Model

MAI-Thinking-1 is the headline announcement. It is a 35-billion active parameter reasoning model using a sparse Mixture of Experts architecture, which means it draws on a much larger pool of knowledge (roughly one trillion total parameters) while keeping inference costs lower than a dense model of the same scale. The context window is 256,000 tokens, enough to process a 600-page document in one pass.

The benchmark numbers are strong. On AIME 2025, the world's most demanding public mathematics competition, the model scored 97.0%. On AIME 2026 (a harder target because 2026 competition problems are far less likely to appear in training data), it scored 94.5%. On SWE-Bench Pro, a software engineering benchmark, Microsoft says it matches Claude Opus 4.6 on coding tasks.

The most commercially relevant number is from the human preference test. Microsoft partnered with Surge, an independent human rating firm, and ran 1,276 single-turn and multi-turn tasks. Professional raters preferred MAI-Thinking-1 over Claude Sonnet 4.6 in blind evaluations where they could not see which model produced each response.

For enterprise buyers: McKinsey fine-tuned the model on their own private data (a feature Microsoft calls Frontier Tuning), and the resulting model outperformed OpenAI's GPT-5.5 at ten times better cost efficiency. Any company can now do the same thing through Microsoft Foundry.

What MAI-Thinking-1 Means for Existing Tools

If you are paying for Claude Pro ($20/mo), ChatGPT Plus ($20/mo), or Gemini Advanced ($19.99/mo) as a general reasoning assistant, MAI-Thinking-1 is the first credible reason to reconsider. It is not available as a standalone consumer subscription yet, but it is accessible through Azure Foundry and will power Microsoft 365 Copilot features.

For developers building on top of AI APIs, this is the most important competitive entrant since DeepSeek V3 forced pricing down across the industry. Another strong model at lower inference cost means more negotiating leverage and more options.

MAI-Code-1-Flash: The Tiny Coding Model Already Inside GitHub Copilot

MAI-Code-1-Flash has 5 billion parameters, a fraction of frontier model sizes, yet it outperforms Claude Haiku 4.5 across all four core coding benchmarks Microsoft tested. The gap on SWE-Bench Pro is not close: 51.2% for MAI-Code-1-Flash versus 35.2% for Haiku 4.5, a 16-point lead. On instruction following (IF Bench), the lead is 28.9 points.

Microsoft did not benchmark it externally and then deploy it. They trained it directly inside GitHub Copilot's production harness, which is why its real-world performance holds up rather than degrading from benchmark to deployment.

The model is rolling out now across every GitHub Copilot tier including Free, Pro, Pro+, and Max, starting with a limited set of users and expanding over the coming weeks. If you already pay for GitHub Copilot, this is a free upgrade you will receive automatically.

MAI-Transcribe-1.5: Faster Than Deepgram and Google

MAI-Transcribe-1.5 covers 43 languages and Microsoft claims it is the most accurate transcription model available, beating Gemini and OpenAI's Whisper on Microsoft's own benchmarks. The official keynote states it runs up to five times faster than rival models.

It is going into Microsoft Teams, Copilot, GitHub, and Dynamics 365 Contact Centre. Developers can access it through Azure Foundry.

If you are paying for Otter.ai ($16.99/mo) or Fireflies ($18/mo) as a dedicated transcription tool and your team runs on Microsoft 365, check whether Teams Copilot covers your use case before renewing.

MAI-Voice-2: Microsoft's Answer to ElevenLabs

MAI-Voice-2 generates speech in 15 languages with fine-grained emotional control and native-sounding delivery. It can clone a voice from a short audio sample. All outputs are watermarked, and Microsoft says the model includes protections against unauthorized voice cloning.

There is also a companion model called Voice-2-Flash, designed specifically for real-time voice agents where ultra-low latency is the priority.

This is a direct challenge to ElevenLabs ($5/mo Starter, $22/mo Creator), which currently leads on voice realism and cloning quality. ElevenLabs supports 32 languages versus MAI-Voice-2's 15 for now, but the gap will close. Microsoft's advantage is distribution: MAI-Voice-2 will be embedded in every Microsoft 365 product without any additional subscription.

MAI-Image-2.5: Second Place in the World for Image Generation

MAI-Image-2.5 ranked second on the Arena-Score image generation benchmark, behind OpenAI's GPT-Image-2 and ahead of Google's image models. It handles both text-to-image generation and image editing.

It is already live inside PowerPoint (for slide image generation) and OneDrive (for photo cleanup). Microsoft 365 subscribers will encounter it without needing to visit any separate tool.

For standalone image generation, Midjourney ($10/mo) still leads on artistic quality for creative work. But for business users who just need images inside presentations and documents, MAI-Image-2.5 makes Adobe Firefly ($10/mo) harder to justify as a separate purchase.

Scout: The Always-On Microsoft 365 Agent

Scout is Microsoft's always-on AI agent for Microsoft 365. It monitors your calendar, reschedules meetings when conflicts arise, guards your deadlines, and manages your inbox while you are offline. It runs across your email, calendar, Teams, and documents, with access to your Microsoft Graph data.

Scout is coming to Microsoft 365 subscribers this summer.

For context: Reclaim AI ($12/mo) and Motion ($29/mo) built their entire businesses around calendar optimization and AI scheduling. Scout does the same thing, bundled into a subscription 60 million enterprise users already have. If you pay for either of these tools and your team uses Microsoft 365, evaluate Scout before renewing.

Maia 200: Microsoft's Own AI Chip

Every model in the MAI family runs on Microsoft's Maia 200 chip, built specifically to run its own models in Azure data centers. The chip delivers 10 petaFLOPS of compute at 750 watts, manufactured on TSMC's 3nm process.

Microsoft's EVP of Cloud and AI stated the chip is 30% cheaper than any competing AI silicon currently available. Running MAI models on the Maia 200 end to end delivers a 1.4x improvement in performance per watt compared to Nvidia's GB200.

When the compute underneath every Microsoft AI product is 30% cheaper to run, that cost advantage compounds across every MAI model, every Copilot feature, and every Azure AI API call. It also reduces Microsoft's single largest external dependency: its GPU spend with Nvidia.

The Bigger Picture: What Microsoft Just Changed

Before Build 2026, Microsoft's AI business ran on two dependencies: OpenAI's models and Nvidia's chips. After Build 2026, Microsoft has its own reasoning model, its own coding model, its own transcription model, its own voice model, its own image model, and its own chip.

For users of third-party AI tools, the practical implication is straightforward. If you pay for a Microsoft 365 subscription and a separate AI voice tool, transcription tool, or image generation tool, check your Microsoft 365 benefits before your next renewal. Several of those categories are now covered at no extra cost.

For developers, MAI-Thinking-1 and MAI-Code-1-Flash are now options in Azure Foundry alongside GPT-5.5, Claude, Gemini, and others. More competition at this level consistently produces lower prices and better models. That is good for everyone building on AI.

Source: All benchmark numbers from official Microsoft Build 2026 keynote transcript and Microsoft AI news releases, June 2 to 3, 2026.