If you use GitHub Copilot, you may already be running Microsoft’s first in-house coding model without knowing it. On June 2 at Build 2026, Microsoft rolled out MAI-Code-1-Flash to all Copilot plans as part of a broader launch of seven self-built AI models. This is not a repackaged OpenAI model. It is a 5-billion-parameter model trained directly on Copilot production workflows, optimized for the tasks that dominate a developer’s day: inline edits, quick refactors, and single-file fixes.
MAI-Code-1-Flash: What It Is and What It Actually Handles
The defining choice behind MAI-Code-1-Flash is where it was trained. Rather than building against synthetic benchmark suites, Microsoft trained it using the actual Copilot production harnesses including the multi-step file editing flows, inline chat sessions, terminal calls, and context retrieval patterns that real users generate every day. The model learned to be good at Copilot, not just good at approximating Copilot.
That distinction shows up in the numbers. On SWE-Bench Pro, MAI-Code-1-Flash scores 51.2% against Claude Haiku at 35.2%, a 16-point gap. On SWE-Bench Verified, it hits 71.6% versus 66.6%. It also uses up to 60% fewer tokens on harder coding tasks, which means faster responses. The mechanism is what Microsoft calls adaptive solution length: the model adjusts how much reasoning it applies based on task complexity. A variable rename uses minimal budget; a multi-file refactor gets the full allocation.
A caveat worth naming: these benchmarks measure exactly what the model was trained to excel at. That is specialization, not cheating, but it means the numbers will look better than what you would see on tasks outside the Copilot harness. Independent third-party replication is not yet available. Real-world signal will emerge from usage data over the coming weeks.
To use it now: open the GitHub Copilot Chat model picker in VS Code and select MAI-Code-1-Flash, or leave the Auto picker on. Copilot may already be routing your inline tasks through it automatically.
MAI-Thinking-1: The Heavy Lifter (Private Preview)
The second model worth knowing is MAI-Thinking-1, Microsoft’s first in-house reasoning model. It runs on a sparse Mixture-of-Experts architecture with 35 billion active parameters and roughly 1 trillion total, with a 256,000-token context window. It was trained from scratch on commercially licensed data with no OpenAI data and no distillation from third-party models.
On benchmarks: 97.0 percent on AIME 2025 and 94.5 percent on AIME 2026. In a blind human evaluation of 1,276 tasks run by rating partner Surge, raters preferred MAI-Thinking-1 over Claude Sonnet. On SWE-Bench Pro, Microsoft claims parity with Claude Opus. These are vendor-reported figures; independent replication is pending.
MAI-Thinking-1 is currently in private preview on Microsoft Foundry, with Baseten among the early hosting partners. Most developers will not have access for months. The target use cases when it opens: complex multi-step reasoning, long-context code generation, and enterprise agentic workflows.
The Bigger Picture: Microsoft Plan B
The full MAI family spans seven models: reasoning, coding, image generation in two variants, transcription, and voice synthesis in two variants. The stated goal is long-term self-sufficiency. The more accurate read is optionality.
OpenAI still accounts for 45 percent of Microsoft cloud backlog, and GPT-5.4 remains the primary model powering most Copilot features. The MAI launch does not change that immediately. What it does is give Microsoft a credible alternative and signals to the market that the company is no longer solely dependent on its largest AI partner. As CNBC noted, this is Microsoft making dependence look optional. That distinction between optionality and full independence is what actually shifted at Build 2026.
For developers, the immediate action is straightforward: check your Copilot model picker in VS Code. MAI-Code-1-Flash is already rolling out. If you want early access to MAI-Thinking-1 when it opens, the Microsoft Foundry waitlist is the starting point.
