Microsoft's MAI-Code-1-Flash: The First AI Microsoft Built Without OpenAI

By crayfish · June 07, 2026 · Category: AI Tools

The Model Microsoft Built Itself

On June 2, 2026, at its annual Build conference, Microsoft unveiled something that would have been unthinkable just two years ago: a frontier AI model built entirely in-house, without a single line of OpenAI technology inside it. The model is called MAI-Code-1-Flash, and it represents a strategic shift that will reverberate across the entire AI industry.

For years, Microsoft’s AI strategy was synonymous with its partnership with OpenAI. The company invested $13 billion into OpenAI, integrated GPT models into every product from Word to Windows, and built its AI identity around that relationship. But MAI-Code-1-Flash signals that Microsoft is no longer content to be OpenAI’s distribution partner. It wants to be a first-class AI model builder in its own right.

The results speak for themselves. MAI-Code-1-Flash outperforms Claude Haiku 4.5 — widely regarded as the best small coding model available — on every major benchmark. And it does so while using significantly fewer tokens, at lower latency, and with better agentic behavior. This is not a research demo. It is a production model, available today inside GitHub Copilot.

Architecture: Small but Mighty

MAI-Code-1-Flash uses a sparse Mixture-of-Experts (MoE) architecture with 137 billion total parameters but only 5 billion active parameters per inference pass. This design philosophy — large knowledge capacity, small computational footprint — is what makes the model both powerful and fast.

The 5 billion active parameter count is deceptively modest. During each inference, the model routes the input through a gating network that selects the most relevant expert subnetworks from the full 137 billion parameter pool. The result is a model that carries the knowledge of a much larger system but runs with the speed and cost of a small one.

One detail that sets MAI-Code-1-Flash apart from virtually every other frontier model is its training data. Microsoft trained MAI-Code-1-Flash exclusively on commercially licensed data. No scraped code from public repositories without consent. No ambiguous licensing situations. This is a model that enterprises can deploy without worrying about copyright exposure — a significant selling point in an industry increasingly haunted by legal questions around training data.

The architecture also delivers a concrete practical advantage: MAI-Code-1-Flash uses 60% fewer tokens than comparable models when solving difficult coding tasks. Fewer tokens means lower cost, lower latency, and faster time-to-answer. In a production coding environment where developers are waiting for AI responses dozens of times per day, that efficiency compounds rapidly.

Benchmark Results: Beating Claude Haiku 4.5

The benchmark numbers are where MAI-Code-1-Flash makes its strongest statement. Across three major coding benchmarks, it consistently outperforms Claude Haiku 4.5, which until now was considered the gold standard for small, fast coding models.

SWE-Bench Verified: MAI-Code-1-Flash scores 71.6%, compared to Claude Haiku 4.5’s 66.6%. That is a +5.0 percentage point improvement, representing a meaningful jump in real-world software engineering task completion.

SWE-Bench Pro: The gap widens significantly on the harder benchmark. MAI-Code-1-Flash achieves 51.2% versus Claude Haiku 4.5’s 35.2% — a staggering +16.0 percentage point lead. On complex, multi-file software engineering challenges, MAI-Code-1-Flash is in a different league.

Terminal Bench 2: For command-line and terminal-based coding tasks, MAI-Code-1-Flash scores 54.8% compared to Claude Haiku 4.5’s 41.6%, a +13.2 percentage point advantage. This matters because terminal interaction is a core part of the developer workflow, and a model that excels here delivers immediate practical value.

Beyond raw scores, developers who have used MAI-Code-1-Flash report three qualitative advantages. First, lower latency — responses come back noticeably faster, which matters when you are in a flow state. Second, better agentic behavior — the model knows when to stop, when to ask for clarification, and when to iterate. Third, more concise outputs — the model tends to produce shorter, more targeted responses rather than over-explaining.

How to Use It Today

MAI-Code-1-Flash is not a waitlist product or a limited beta. It is available right now inside GitHub Copilot, and enabling it takes less than two minutes.

Step 1: Update VS Code to version 1.99 or later. The latest release includes the Copilot model selector that supports MAI models.

Step 2: Open the Copilot Chat panel in VS Code. If you do not see it, press Ctrl+Shift+P and search for “GitHub Copilot: Chat.”

Step 3: Click the model selector dropdown at the top of the Copilot Chat panel. You will see “MAI Code 1 Flash” listed alongside the standard GPT models. Select it.

That is it. From this point forward, your Copilot interactions will use MAI-Code-1-Flash by default. You can switch back to GPT-based models at any time through the same selector, allowing you to compare outputs side by side.

For developers who prefer the API route, MAI-Code-1-Flash is available through the Azure AI endpoint with straightforward pricing: $0.75 per million input tokens and $4.50 per million output tokens. At 60% fewer tokens per task compared to competitors, the effective cost advantage is even larger than the per-token pricing suggests.

The Growing MAI Family

MAI-Code-1-Flash is not a one-off experiment. It is part of a broader model family that Microsoft is building under the MAI (Microsoft AI) brand. The family already includes several specialized models:

• MAI-Thinking-1: A 35 billion parameter reasoning model designed for complex analytical tasks, mathematical problem-solving, and multi-step logical reasoning.
• MAI-Code-1-Flash: The coding specialist described in this article, optimized for software engineering tasks with production-grade performance.
• MAI-Transcribe-1.5: A speech-to-text model with improved accuracy across accents, languages, and noisy environments.
• MAI-Voice-2: A text-to-speech model with natural prosody and emotional range for conversational applications.
• MAI-Image-2.5: An image generation and understanding model for visual AI tasks.

The message is clear: Microsoft is building a full-stack AI model portfolio that covers every major modality — text, code, voice, and image. Some of these models will compete with OpenAI’s offerings. Others will complement them. But the strategic independence they represent is the real story.

What This Means for the AI Landscape

Microsoft’s decision to build its own frontier models is one of the most significant strategic shifts in the AI industry this year. It does not mean the OpenAI partnership is ending — Microsoft has publicly stated that it remains committed to the relationship. But it does mean that Microsoft now has options, and options change power dynamics.

For developers, MAI-Code-1-Flash is simply a better tool for certain tasks. The benchmark numbers are real, the latency improvement is noticeable, and the commercially clean training data removes a category of enterprise concern entirely. For the broader industry, it signals that the era of single-vendor AI dependency may be ending. Companies like Microsoft, Google, Meta, and Anthropic are all building their own model stacks, and the competition is producing better models at a pace that benefits everyone.

The question is no longer whether Microsoft can build AI models without OpenAI. MAI-Code-1-Flash answers that question definitively. The question now is how far Microsoft will take this capability — and what it means for the companies that once had the field to themselves.