Microsoft has put a new pair of in-house coding models into GitHub Copilot: MAI Code 2 for heavier repository work and MAI Code 2 Flash for quicker, repeatable tasks. The launch matters because it gives developers another selectable option inside a familiar tool. The harder question is whether Microsoft’s attractive benchmark and price comparisons predict results on a team’s own code.
The models were announced on November 17, 2026, alongside Microsoft Ignite. GitHub says both are available through Copilot’s model picker on supported clients, including Visual Studio Code, Visual Studio, the CLI, JetBrains, Xcode and github.com. Eligibility depends on subscription: Code 2 is listed for Pro, Pro+, Max, Business and Enterprise plans; Flash is also listed for Free and Student. GitHub says use is billed under usage-based billing, so access does not mean unlimited included use. See GitHub’s availability and billing details.
Two models for different kinds of work
Microsoft positions Code 2 for debugging, multi-file changes, refactoring and reasoning across a repository. Flash is aimed at code explanations, narrow edits, routine fixes and quick iterations. In principle, the split lets a developer reserve a larger model for tasks where a wrong edit would cost more review time and use the smaller one for frequent, bounded requests. That is a product strategy, not evidence that every task will route to the best model automatically; developers still choose from Copilot’s model picker.
The models are hosted services, not downloads for a developer’s own machine. Microsoft’s model card says the weights are not distributed and the models run on Microsoft Azure GPU infrastructure. Teams that need local execution or control over model weights should not read this launch as providing either.
What Microsoft’s benchmark claims do—and do not—show
Microsoft reports that Code 2 averaged 98% of GPT 6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s performance at 5% of the cost, across four coding benchmarks. Those percentages are Microsoft’s results, not an independent head-to-head study of every repository or workflow. Microsoft says three benchmarks ran on its internal MSBench platform using Copilot CLI, at maximum reasoning effort and without subagents; the fourth used Artificial Analysis results. The cost comparison also uses a 50% promotional price for MAI Code 2.
There is a useful boundary in the announcement: Microsoft says the models are not yet tuned for fundamental science research or very long-horizon agent work. That makes the reported average most relevant to the everyday enterprise coding tasks Microsoft selected. It should not be treated as a general ranking for all software engineering or as a forecast of how much an individual team will save.
A practical evaluation should replay a small, representative set of real tasks: bug fixes, multi-file changes, test repair and repository questions. Keep the prompt and repository state the same, record whether the model’s changes pass tests, and measure reviewer corrections and elapsed time as well as token cost. A cheaper attempt that needs multiple rounds of repair may not be cheaper in practice.
The model card is part of the story
Microsoft’s published model card says Code 2 and Flash can repeat tool calls, underuse existing code, drift from instructions or stop before fully checking their work during long sessions. It also says performance outside English and on less common languages and frameworks has not been extensively evaluated. Microsoft advises against deploying generated code to production or safety-critical systems without human review and testing. These caveats describe exactly why repository-specific trials and ordinary code review remain necessary.
For engineering managers, the near-term decision is modest: whether to enable another model option and how to test its quality, latency, usage charges and policy fit. Business and Enterprise administrators can manage access through Copilot’s model policy, according to GitHub. A team that tries it should set a usage budget, choose low-risk tasks first and compare reviewed outcomes against its current workflow before making it a default.
Sources
Microsoft AI: MAI Code 2 and MAI Code 2 Flash announcement · GitHub Changelog: availability in Copilot · Microsoft: MAI Code 2 model card
