
Microsoft AI released MAI Code 2 and MAI Code 2 Flash on November 17, 2026 and put them in GitHub Copilot. Every performance number so far is Microsoft's own.
Microsoft AI released MAI Code 2 and MAI Code 2 Flash on November 17, 2026, the Tuesday of its Ignite conference, and moved both into GitHub Copilot the same day. The headline claim is a price-performance one: 98 percent of GPT-6.1 Sol's score at 40 percent of the cost. It is a striking number, and every input to it belongs to Microsoft.
What shipped
MAI Code 2 targets what Microsoft calls common enterprise coding tasks. MAI Code 2 Flash targets high-volume routine work: explanations, targeted edits and quick fixes. Microsoft is candid about the limit. In its words, the models are "explicitly not yet tuned for harder problems, like fundamental science research," or very long-horizon agent work.
GitHub's changelog, dated November 17, 2026, says MAI Code 2 is available to Copilot Pro, Pro+, Max, Business and Enterprise users, and Flash additionally to Free and Student users, on annual plans where eligible. Both are billed under usage-based billing, and the changelog sends readers to its pricing documentation rather than listing prices. Business and Enterprise administrators control access through the model policy, and new models are enabled by default unless an admin switches the default off. Microsoft says Flash is also in limited-access preview in Microsoft Foundry, with OpenRouter, Vercel, Hermes, OpenCode and Kilo listed as coming soon.
The claims, line by line
Microsoft averages four benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average it says:
- MAI Code 2 delivers 98 percent of GPT-6.1 Sol's performance at 40 percent of the cost.
- It delivers 96 percent of Claude Opus 5.5's performance at 5 percent of the cost.
- In blind frontend evaluations with 258 annotators, developers preferred it to GPT-6 Sol 64 percent of the time.
- MAI Code 2 Flash is comparable to GPT-6 Luna and Claude Haiku 5.5 at 65 percent and 20 percent of their cost on three routine-coding benchmarks.

Why the numbers need care
First, the comparison set is not the current frontier. Microsoft compares against GPT-6.1 Sol and GPT-6 Sol, while OpenAI shipped GPT-6.2 Astra and Sol on November 17, the same day, as we reported in our piece on the GPT-6.2 launch. It compares against Claude Opus 5.5, while Anthropic released Fable 5.2 and Mythos 5.2 this month. A model that reaches 96 percent of an older flagship is not reaching 96 percent of the best available one.
“They're explicitly not yet tuned for harder problems, like fundamental science research”
Second, the cost side of the ratio is built on a promotion. Microsoft says costs are shown at a 50 percent promotional price. A cost ratio that halves when the promotion ends tells you about a launch discount, not a steady-state price.
Third, an average across four benchmarks hides spread. A model can score near a rival on average while losing badly on the tasks you care about. Microsoft's post publishes the averages, and we could not find the underlying prompts or scoring details in it. No independent evaluator had published a score for either model when we checked on November 19.
What it does show
It would be wrong to dismiss the release. Microsoft has now put its own coding model in the world's most-used developer tool, next to models from the labs it also partners with. That is a real strategic move: Microsoft AI is building models it controls end to end, and pricing them to compete with partners' flagships. The company also released MAI Work Flash for Microsoft 365, tuned from Code 2 Flash, claiming parity with GPT-6 Luna on Excel and PowerPoint tasks and about 42 percent fewer input tokens per call than GPT-6.1 Sol on Word tasks.
For the race, the signal is the ladder. Every lab with a platform is building a cheap tier that is good enough for routine work, and the premium tier is shrinking to the hardest problems. Microsoft's own caveat, that these models are not tuned for research-grade tasks, says the same thing.
What an annotator study can and cannot tell you
The blind frontend result deserves a second look because it is the one number that is not a benchmark average. Microsoft says 258 annotators preferred MAI Code 2 to GPT-6 Sol 64 percent of the time, which works out to roughly 165 preference votes if each annotator voted once. That is a reasonable sample for a rough preference, and a poor one for fine ranking. Preference tests measure how appealing an interface looks and feels, not whether the code survives review, and the post does not say how tasks were chosen or whether annotators saw the same prompts. It also compares against GPT-6 Sol, not the GPT-6.2 Sol that OpenAI released the same day.
A practical reading follows. Use the model for tickets where review catches mistakes cheaply, such as formatting changes, test scaffolding and small refactors, and keep stronger models for changes that touch security or data. Admins who want to control spend can use the Copilot model policy that GitHub describes in its changelog.
Our take
Take the release as a credible, cheap option that Copilot users can switch on today, not as proof that a smaller lab has matched the frontier. A fair test is easy: pick ten real tickets from your own backlog, run them through MAI Code 2, your current default and one frontier model, and compare cost per accepted change. Until an independent score appears, 98 percent is a marketing figure with an honest caveat attached.
Frequently asked questions
What is MAI Code 2?
MAI Code 2 is a coding model from Microsoft AI released November 17, 2026, aimed at common enterprise coding tasks. It sits alongside MAI Code 2 Flash, a cheaper model for high-volume routine work.
Is MAI Code 2 as good as GPT-6.1 Sol?
Microsoft says it delivers 98 percent of GPT-6.1 Sol's average score across four coding benchmarks at 40 percent of the cost. These are Microsoft's own figures and no independent replication was published as of November 19, 2026.
Where can I use MAI Code 2?
GitHub says MAI Code 2 is available in Copilot Pro, Pro+, Max, Business and Enterprise, and Flash also in Free and Student. Flash is in limited preview in Microsoft Foundry.
How much does MAI Code 2 cost?
GitHub's changelog says the models are billed under usage-based billing and points to its pricing documentation. Microsoft's post shows costs at a 50 percent promotional price, so the later price may be higher.
What is MAI Code 2 not good at?
Microsoft says the models are not yet tuned for harder problems such as fundamental science research or very long-horizon agentic work.
Sources
What each one is, and whose it is.
- 1
Introducing MAI Code 2 and MAI Code 2 Flash, Microsoft AI (November 17, 2026)
Vendor announcement - 2
MAI Code 2 and MAI Code 2 Flash in GitHub Copilot, GitHub (November 17, 2026)
Documentation - 3
Everything that happened in AI today, Wednesday, November 18, The Neuron (November 18, 2026)
Press reportIndependent of the vendor