Short answer: MAI Code 2 is Microsoft's new coding model, built by its Microsoft AI (MAI) team for everyday software work like debugging, multi-file changes and UI code. Microsoft released it on November 17, 2026 at Ignite, together with a lighter sibling, MAI Code 2 Flash (Microsoft AI). Both are already in the GitHub Copilot model picker. MAI Code 2 is on paid Copilot plans, and Flash also reaches Copilot Free and Student (GitHub Changelog). Microsoft says it gets close to bigger models for much less money, but it also says the model isn't tuned yet for science or very long agent tasks.
The announcement had more than 400 points on Hacker News by November 18, and Copilot users will see the name whether they go looking or not, because the older MAI-Code-1.1-Flash leaves Copilot in December. This guide covers what the model is, how Code 2 and Flash differ, which plan gets which, what Microsoft's numbers really say, and when another model is still the better pick.
What is MAI Code 2?
When you pick a model in Copilot, you're choosing the AI engine that reads your code and writes the changes. Until recently, most of those engines came from OpenAI, Anthropic or Google. MAI Code 2 is Microsoft's own.
Microsoft describes it as a model for "everyday enterprise coding tasks." Its list of strengths is practical: taking code to production, deep debugging, changes that span several files, reasoning across a whole repository, and user interfaces. Microsoft also says it's tuned for dashboards and data visualizations.
The pitch is price. Microsoft built the model around the coding work it sees most in GitHub Copilot, and aimed for good results at a fraction of what frontier models cost.
MAI Code 2 vs MAI Code 2 Flash
| MAI Code 2 | MAI Code 2 Flash | |
|---|---|---|
| Built for | Deep debugging, multi-file work, refactoring, repo-wide reasoning | Code explanations, targeted edits, routine fixes, fast iteration |
| Copilot plans | Pro, Pro+, Max, Business, Enterprise | Free, Student, Pro, Pro+, Max, Business, Enterprise |
| Think of it as | The one you pick for a real task | The one you leave on for quick questions |
GitHub notes that both models are for eligible subscribers on annual plans and are billed under usage-based billing, so check how your plan counts usage before you switch everything to them. On Business and Enterprise, admins control access through the model policy in Copilot settings. New models are on by default unless an admin turned that off.
The Flash detail matters if you're learning. Copilot Free and Student users now get a current coding model in the same picker as paid users.
Where can you use it?
GitHub lists the model picker in Visual Studio Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, the Copilot cloud agent, the GitHub Copilot app, github.com and GitHub Mobile.
Outside Copilot, Microsoft says MAI Code 2 Flash is in a limited access preview in Microsoft Foundry. It also says both models are coming "soon" to OpenRouter, Vercel, Hermes, OpenCode and Kilo. There's no date for those yet.
If you use MAI-Code-1.1-Flash today, GitHub will retire it across Copilot on December 18, 2026 and suggests the latest MAI Code Flash model instead (GitHub Changelog).
What do Microsoft's numbers say?
Microsoft's headline claim: averaged over four coding benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), MAI Code 2 gets 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.
Read those numbers with three things in mind:
- They're Microsoft's. Three of the four benchmarks ran on Microsoft's internal evaluation platform with the Copilot CLI. FrontierCode scores and costs come from Artificial Analysis.
- The cost uses a promotional price. Microsoft's footnote says MAI Code 2's cost is shown at its 50% promotional price. When the promotion ends, the gap narrows.
- The comparison is already a step behind. OpenAI released GPT-6.2 Sol on the same day, and Microsoft compared against GPT-6.1 Sol.
Microsoft also says developers preferred MAI Code 2's frontend results over GPT-6 Sol's 64% of the time in blind side by side tests. For Flash, it claims performance comparable to GPT-6 Luna and Claude Haiku 5.5 at a lower cost.
These are reasonable starting points, not a verdict. The only benchmark that counts is your own repository.
Where MAI Code 2 is weaker
Microsoft is unusually direct about this. It says both models are "explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work."
In practice, that means two kinds of work:
- Long agent runs. If you hand an agent a big task and walk away for an hour, while it plans, edits, runs tests and tries again, Microsoft is telling you this isn't the model's strength yet.
- Research heavy code. Scientific computing and problems where the hard part is the math, not the code.
When should you pick GPT-6.2 Sol or Claude instead?
A simple way to decide:
- Pick MAI Code 2 for everyday tasks you'd do yourself in an afternoon: a bug in one feature, a refactor across a few files, a dashboard or a UI change. That's where Microsoft tuned it, and where the lower cost helps most.
- Pick MAI Code 2 Flash for quick questions, explanations and small edits, especially on Copilot Free or Student.
- Pick GPT-6.2 Sol if you want OpenAI's newest everyday model and it's available where you work. OpenAI pitches it for reviewing pull requests and revising code after feedback (OpenAI), and Microsoft's comparison doesn't cover it.
- Pick a frontier model, such as Claude Opus 5.5 or OpenAI's GPT-6.2 Astra, for long autonomous agent runs or science heavy problems, the two areas Microsoft says MAI Code 2 isn't tuned for yet.
Whatever you pick, try the same real task on two models before you switch your default. It takes ten minutes and tells you more than any benchmark table.
Go deeper
- Best AI coding agents: the main coding tools of 2026, compared.
- What is GPT-6: OpenAI's model names, including GPT-6.2 Astra and Sol.
- Claude Haiku vs Sonnet: when the small model is the right call.
- Microsoft Copilot Home, Code and Autopilot: which Copilot mode fits which job.
- AI tools for developers: the 2026 tools, sorted by what they do.
Picking models is a skill
New coding models now ship every few weeks. Knowing how to test one on your own code, read a vendor's benchmark footnotes and choose the right model for each task is worth more than any single model. If you want to build that skill at your own pace, see how AI Flex works.
