Microsoft MAI Code 2: Near-Frontier Coding at 5% of the Cost?
Microsoft launched MAI Code 2 and MAI Code 2 Flash at Ignite on November 17, 2026, and both are generally available now in GitHub Copilot. Microsoft claims near-frontier coding performance at a fraction of the cost, with a 50% promotional price in the fine print.
AI Stella and Liam are AI hosts with AI-generated voices. About our AI hosts
Key takeaways
- Microsoft launched MAI Code 2 and MAI Code 2 Flash at Ignite on November 17, 2026. Code 2 is built for everyday engineering like debugging and multi file changes; Flash is aimed at high volume jobs like explanations, targeted edits, and routine fixes.
- Microsoft claims Code 2 reaches 98% of GPT-6.1 Sol performance at 40% of the cost, and 96% of Claude Opus 5.5 performance at 5% of the cost, averaged over selected coding evaluations at maximum reasoning effort.
- The fine print matters: three tests ran on Microsoft's internal platform through the GitHub Copilot CLI without subagents, the figures are vendor reported and not independently reproduced, and the MAI cost math uses a 50% promotional price.
- Trying it takes one click: both models are generally available in the GitHub Copilot model picker, and Code 2 Flash is also in limited preview in Microsoft Foundry.
- Microsoft says the models are not yet tuned for harder problems, including fundamental science and very long running agent work, so keep a frontier model for the hardest jobs.
Show notes
Microsoft launched MAI Code 2 and MAI Code 2 Flash at Ignite on November 17, 2026, and both are generally available now in GitHub Copilot. Microsoft claims Code 2 reaches 98% of GPT-6.1 Sol performance at 40% of the cost, and 96% of Claude Opus 5.5 performance at 5% of the cost, on selected coding evaluations at maximum reasoning effort. The MAI cost math uses a 50% promotional price, and Microsoft says the models are not yet tuned for harder problems, including very long running agent work. In this episode of Agentic Pod, Liam and Stella explain what shipped, read the fine print, show how to try it in the Copilot model picker, and give an honest verdict.
What we cover:
- What shipped: MAI Code 2 for everyday engineering work like debugging, multi file changes and user interface work, and Code 2 Flash for high volume jobs like explanations, targeted edits and routine fixes
- Availability: generally available in GitHub Copilot, with Code 2 Flash also in limited preview in Microsoft Foundry
- The headline numbers: 98% of GPT-6.1 Sol performance at 40% of the cost, and 96% of Claude Opus 5.5 performance at 5% of the cost, per Microsoft’s own selected evaluations
- The fine print: selected coding evaluations at maximum reasoning effort, three tests on Microsoft’s internal platform through the GitHub Copilot CLI without subagents, and MAI costs calculated at a 50% promotional price
- How to try it step by step: open GitHub Copilot, open the model picker, pick Code 2 or Code 2 Flash, and test it on a real task from your own work
- A real workflow: Flash for the pile of small fixes, Code 2 for debugging across files, and a frontier model saved for the hardest problem
- Gotchas: not yet tuned for harder problems or very long running agent work, promo pricing, and vendor run benchmarks
Chapters
Questions answered in this episode
What is MAI Code 2?
Microsoft's own coding model, launched at Ignite on November 17, 2026. It is built for everyday engineering work like debugging, multi file changes, and user interface work, and it is generally available in GitHub Copilot.
What is MAI Code 2 Flash?
The smaller, faster model in the same launch, aimed at high volume tasks like explanations, targeted edits, and routine fixes. It is in GitHub Copilot and in limited preview in Microsoft Foundry.
Should I trust the 98% and 96% numbers?
They are Microsoft's own figures, averaged over selected coding evaluations at maximum reasoning effort, with three tests on Microsoft's internal platform, and the cost side uses a 50% promotional price. Treat them as a reason to test the models on your own work, not as a settled result.
How do I try MAI Code 2?
Open GitHub Copilot, open the model picker, pick Code 2 or Code 2 Flash, and give it a real task from your own work. Compare the result and the bill against your usual model.
Links & resources
Full transcript
Intro
Hey hey hey, welcome to Agentic Pod! I'm Liam.
And I'm Stella! Liam, Microsoft launched MAI Code 2 at Ignite on November 17th, and the headline is a price claim?
A big one. Microsoft says Code 2 reaches 98 percent of GPT-6.1 Sol performance at 40 percent of the cost.
And against Claude?
96 percent of Claude Opus 5.5 performance at just 5 percent of the cost. By the end you will know what is real, what is fine print, and how to try it today.
What Microsoft actually shipped
Okay, plain words. What is MAI Code 2?
It is Microsoft's own coding model. It is made for everyday engineering work, like debugging, changes across files, and user interface work.
And there is a smaller one too?
Code 2 Flash. That one is aimed at high volume jobs. Explanations, targeted edits, and routine fixes.
So one model for the deep work, one for the quick jobs.
Exactly. And both are generally available now inside GitHub Copilot. Flash is also in a limited preview inside Microsoft Foundry.
Wait, it is already in Copilot? No waitlist, no new tool to install?
No new tool. It sits in the model picker, next to the models you already use. Which raises the obvious question, how much should you trust those headline numbers?
The numbers, and the fine print
Right. 98 percent and 96 percent sound huge. What did Microsoft actually measure?
Selected coding evaluations, averaged together. The models ran at maximum reasoning effort, and three of the tests ran on Microsoft's own internal platform, through the GitHub Copilot CLI, without subagents.
So it is Microsoft's test, on Microsoft's setup.
Yes, and nobody outside Microsoft has reproduced it yet. It does not mean the model finishes 98 percent of your real tasks. It is a relative score on their tests.
And the cost side? That 5 percent number?
Here is the catch. The MAI cost math uses a 50 percent promotional price. So the comparison is real on paper, but it is priced at half price.
Half price, full headline. Classic.
Classic, but still interesting. A cheaper model that needs three retries can cost more than a pricey model that nails it first try. So what does trying it look like?
How to try it, and a real workflow
Step one, open GitHub Copilot, in your editor or the Copilot CLI. Step two, open the model picker.
The same picker where I choose between models today?
The same one. Pick MAI Code 2 for a bigger job, or Code 2 Flash for a quick one. Step three, give it a real task from your own work, not a toy prompt.
Like what?
Like, explain this function and fix the failing test next to it. Then compare. Same task, your usual model, check the result and the bill.
And Foundry?
If your team builds on Microsoft Foundry, Flash is there in limited preview. But for most people, Copilot is the front door. Now, where does this actually fit in a normal day?
Picture a normal morning. You have a bug report, a small feature, and a pile of small fixes waiting.
The pile is always the worst part.
Send the pile to Flash. Explain this code, make this targeted edit, fix this routine bug. Fast, cheap, high volume work.
And the bug report?
That goes to Code 2. Debugging across files is exactly what Microsoft built it for. It reads the repo, follows the change across files, and works on the fix.
So you split the day by job size, not by loyalty to one model.
That is the whole play. Save the expensive frontier model for the one hard problem that truly needs it. Which brings us to the limits, and Microsoft stated them itself...
Gotchas and the honest verdict
Gotchas. Start with the one Microsoft admitted.
Microsoft says these models are not yet tuned for harder problems. That includes fundamental science and very long running agent work.
So no overnight autonomous agent runs on this yet?
Not the pitch, no. Second gotcha, the promo price. The cheap math depends on a 50 percent promotional price, and nobody should budget a whole year on a launch discount.
And third, the benchmarks are Microsoft's own.
Right. Selected tests, max reasoning, internal platform. Promising, but your repo is the only test that counts. Honest verdict, then. Who is this for?
If you live in GitHub Copilot, try it this week. It costs you one model picker click.
Agreed. If you run lots of routine coding work, the Flash split could save real money. If you need the hardest long agent runs, keep your frontier model close. The step by step guide is in the description. Would you switch your daily coding model for a cheaper one at 96 percent? Tell us in the comments. See you next time!