On November 17, 2026, at Microsoft Ignite, Microsoft AI announced three new in-house models. MAI Code 2 and MAI Code 2 Flash are coding models available in GitHub Copilot. MAI Cyber 2 Flash is a security model for verified defenders. Here’s the short verdict.
If you already pay for GitHub Copilot, you may get new, cheaper-sounding options to try, depending on your plan and how far Microsoft has rolled the models out. Everyone else can skip this one. The biggest caveat is that every performance and price comparison comes from Microsoft itself.
Ever feel like a new AI model drops every other Tuesday? You’re not imagining it. This one deserves a closer look because Microsoft is pitching the launch as a lower-cost alternative to named OpenAI and Anthropic models.
What Happened With MAI Code 2 and MAI Cyber 2 Flash
Microsoft AI is the company’s in-house AI division. It introduced the coding pair in its official announcement. The security model has its own MAI Cyber 2 Flash model page. The Ignite 2026 hub and the MAI keynote recap cover the broader launch.
None of these are chatbots you download from an app store. They’re the “engines” that run inside other tools. Think of swapping the motor in a car you already own.
- MAI Code 2: Handles the harder coding jobs. Examples include fixing bugs scattered across many files, building app interfaces, and writing production-ready code. It’s available now in GitHub Copilot, depending on your plan.
- MAI Code 2 Flash: A lighter, faster sibling for quick tasks, such as explaining a chunk of code or making a small, targeted fix. It’s in GitHub Copilot and in limited preview on Microsoft Foundry, Microsoft’s platform for businesses building AI apps.
- MAI Cyber 2 Flash: Built for security teams. Microsoft says it finds vulnerabilities, checks whether they’re real or false alarms, and helps fix them. Access runs only through Microsoft’s Codename MDASH system, and only for verified defenders.
MAI Cyber 2 Flash: Built for Verified Defenders
MAI Cyber 2 Flash targets a narrower audience than the coding models. Microsoft says it matches the quality of OpenAI’s GPT-6 Astra on the CyberGym-E2E security benchmark while costing less than 2% as much per task to run. Its model page also lists a Microsoft-reported 83.2% solved-task score on that benchmark, ahead of GPT-5.6 Cyber, without detailing the evaluation setup.
Access is limited to verified defenders inside Microsoft’s Codename MDASH program, so most readers won’t be able to try it directly. As with the coding comparisons, the CyberGymE2E result is a vendor-reported claim rather than an independently confirmed benchmark run.
How It Stacks Up
Every comparison below is vendor-reported. Microsoft picked the benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, FrontierCode v1.1, and CyberGymE2E.
At publication, the coverage cited below, including Mad Robot and The New Stack, discusses and analyzes Microsoft’s own figures rather than reporting independent benchmark runs of these models.
| Microsoft model | Compared against | Microsoft’s claim (vendor-reported) |
|---|---|---|
| MAI Code 2 | OpenAI GPT-6.1 Sol | About 98% of the coding performance at roughly 40% of the cost |
| MAI Code 2 | Anthropic Claude Opus 5.5 | About 96% of the performance at around 5% of the cost |
| MAI Code 2 Flash | Claude Haiku 5.5 and OpenAI GPT-6 Luna | Comparable to Haiku 5.5 at about 20% of its price, and comparable to GPT-6 Luna at about 65% of its price |
| MAI Cyber 2 Flash | OpenAI GPT-6 Astra | Matches GPT-6 Astra’s quality on the CyberGymE2E benchmark at less than 2% of the cost per task |
These are relative percentages drawn from Microsoft’s own calculations. Microsoft hasn’t published matching dollar prices or confirmed whether the comparisons use the same billing unit, such as per-token or per-task costs, so the figures can’t be translated into an exact cost difference for any specific workload.
Microsoft’s footnotes add another caveat: the MAI Code 2 and Code 2 Flash cost figures are shown at a 50% promotional price, and the announcement doesn’t say how long that promotion lasts.
Mad Robot published its own breakdown of the cost and performance gap behind Microsoft’s percentages. The New Stack framed the launch as “near-Opus performance at 5% of the cost,” which is Microsoft’s pitch in a nutshell.
The honest read is that a model scoring a point or two lower for a fraction of the price is a solid deal for routine work. “Routine” is the key word, though. For the trickiest jobs, those last few percentage points can be the difference between a fix that works and one you redo (and redoing work is never cheap).
The Reaction
Reaction is still forming, and outside coverage so far mostly relays Microsoft’s own performance and cost figures.
Our Take: Who Should Care
GitHub Copilot users: Worth a look, though availability depends on your specific Copilot plan and how far Microsoft has rolled the models out. MAI Code 2 Flash makes sense for quick edits, and MAI Code 2 for bigger jobs, if your plan includes them.
The downside is that Microsoft’s own testing is the only evidence so far. Try the new models on real work before trusting the percentages, and check Copilot’s model picker or Microsoft’s official documentation to confirm what your plan actually offers.
Enterprise security teams: Cyber 2 Flash targets a real pain point, which is chasing false alarms. But you need to qualify as a verified defender through Codename MDASH, so it’s only practical if you already work with Microsoft’s security tools.
Small business owners and everyday readers: These are backend models surfaced through developer tools and security products, not consumer apps you download. Using them still requires an eligible GitHub Copilot plan, Microsoft Foundry preview access, or verified-defender status.
The bigger story is price pressure. Microsoft undercutting OpenAI and Anthropic on these vendor-reported percentages could eventually push AI service costs down for everyone, though Microsoft hasn’t disclosed comparable dollar prices or billing terms, and its coding-model cost figures use a promotional price.
Should you switch coding assistants? Not because of this announcement alone. If you’re happy with Claude or ChatGPT, vendor benchmarks and Microsoft’s pricing claims are thin reasons to move your workflow.
Wrapping Up
Microsoft’s new MAI models are aimed at developers and security pros, not the rest of us. The pricing claims are eye-catching, but they’re Microsoft grading its own homework until independent tests arrive. Use MAI Code 2 or Code 2 Flash if you’re already in GitHub Copilot and want cheaper routine help. Stick with your current assistant if it’s working and you’d rather wait for independent numbers.