2026-11-18 18:21 UTC
DANGMUAAI & Developer Tools, Decoded
BackDev Tools

Microsoft MAI Code 2 Is Live in GitHub Copilot: The Numbers

Microsoft's MAI Code 2 and Code 2 Flash are generally available in GitHub Copilot, with price-performance claims that carry two large caveats.

DangMua EditorialNov 18, 20264 min read
Microsoft MAI Code 2 Is Live in GitHub Copilot: The Numbers

Microsoft put its own coding models into GitHub Copilot on November 17, claiming 96% of Claude Opus 5.5's coding quality at about 5% of the cost.

MAI Code 2 and MAI Code 2 Flash went generally available in Copilot — in the model picker and inside Copilot Fusion — and in Microsoft Foundry, announced at Ignite 2026 in San Francisco as part of a family of seven in-house MAI models. This is Microsoft's first in-house coder shipped as a first-class option inside Copilot, next to the OpenAI and Anthropic models Copilot has historically relied on.

The numbers, and whose numbers they are

Microsoft's headline claims for MAI Code 2 are averages across four coding benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1:

Microsoft modelCompared withClaimed performanceClaimed relative cost
MAI Code 2GPT-6.1 Sol~98%~40% of the cost
MAI Code 2Claude Opus 5.5~96%~5% of the cost
MAI Code 2 FlashGPT-6 LunaComparable65% of the cost
MAI Code 2 FlashClaude Haiku 5.5Comparable20% of the cost

Two caveats come attached to that table, and both are material. All of these figures come from Microsoft's own testing, so they are company claims until independent benchmarks catch up. And the cost ratios use a 50% limited-time promotional price — the gap narrows once the promotion ends. The announcement does not publish dollar-denominated prices or the individual benchmark scores, so there is no way to re-derive the ratios yourself yet.

The Flash figures rest on three day-to-day, high-volume coding benchmarks rather than the four used for Code 2 — a narrower basis for a narrower claim.

What each model is for

Microsoft positions Code 2 for everyday enterprise coding: taking code to production, deep debugging, multi-file implementation, repo-wide reasoning and UI work. Flash covers code explanations, targeted edits, routine fixes and rapid iteration. Per the model card, both use a sparse Mixture-of-Experts architecture; Flash has approximately 12 billion active parameters and a one-million-token context window. Microsoft says Code 2 is larger but has not given a parameter figure.

More useful than either pitch is the scope limit Microsoft volunteered: the models are, in the company's words, "explicitly not yet tuned" for fundamental science research or very long-running agentic tasks. That is an unusually specific exclusion for a launch, and it draws the line cleanly — high-volume production coding, not multi-hour autonomous agent runs.

Why it matters beyond the price tag

The competitive axis has moved. Coding-model launches used to argue capability; this one argues cost-performance on routine work, and it does so from inside the client where the work already happens. With Code 2 sitting in the same picker as Claude and GPT models, and Copilot Fusion routing between them, the operative question for a team is no longer which model is best but which model is good enough for this task at this price — a routing decision, not a procurement one.

It also reduces Microsoft's dependence on a single supplier. An in-house coder in Copilot, plus the newly announced MAI Licensed Weights program giving trusted customers access to Microsoft AI model weights, point the same direction.

What to check before you move work onto it

Three things are worth measuring yourself, since none can be read off the announcement: whether the 96–98% figures hold once independent benchmarks run the same comparisons; how much of the price advantage survives the end of promotional pricing; and whether "everyday enterprise coding" covers enough of your actual work, or whether the hard cases still route to a frontier model. If you use Copilot, Code 2 is already in your model picker — the cheapest test is to send it a week of routine edits and compare the diffs you would have shipped anyway.

More from DangMua