Microsoft used its November 17, 2026, Ignite announcements to pitch its in-house MAI models as specialized alternatives for organizations seeking lower-cost AI for coding, cybersecurity, and other tasks. The headline figures are Microsoft’s own benchmark and cost comparisons: useful for understanding the company’s case, but not evidence that its models perform equally well across all uses.
2
3
10
Why reports say six models and Microsoft says seven
Microsoft described the release as a family of seven models, while Axios reported six specialized models.
2
10 The available materials do not explicitly explain the discrepancy. They do show that the announcements spanned more than coding and cybersecurity: Microsoft and other reporting refer to image, speech, transcription, and decision-making models as well.
5
8
35
That broader lineup may help explain why summaries count the launch differently, but the sources do not establish a definitive reconciliation. The safest description is that Microsoft promoted a seven-model family while Axios counted six in its report.
What Microsoft claimed about coding performance
Microsoft positioned MAI Code 2 for enterprise coding work such as debugging, multi-file changes, taking code to production, and reasoning across a repository.
1 Across four coding benchmarks—DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1—the company said Code 2 averaged 98% of GPT-6.1 Sol’s performance at 40% of its cost, and 96% of Claude Opus 5.5’s performance at 5% of its cost.
3
16
The lighter MAI Code 2 Flash is aimed at higher-volume, more routine work, including code explanations, targeted edits, fixes, and rapid iteration. Microsoft said it performed comparably to GPT-6 Luna at 65% of its price and Claude Haiku 5.5 at 20% of its price.
1
3
These numbers should be read as vendor-reported comparisons on selected coding tasks. Microsoft’s keynote describes the benchmark set as targeting everyday enterprise coding, not every kind of programming or reasoning.
3 They do not establish that either model is a general replacement for its competitors.
Cybersecurity claims and access
For vulnerability discovery, Microsoft reported that MAI Cyber 2 Flash achieved 92.3% task success on CyberGym-E2E, compared with 91.5% for GPT-6 Astra, at less than 2% of the cost per task.
29 Reuters reporting said Microsoft planned to make the security model available initially to a limited group of customers.
32
As with the coding results, the score applies to a particular benchmark. It does not, by itself, show how the model performs across all cybersecurity work or in every customer environment.
Availability and the price caveat
Microsoft’s announcements and related reporting say Code 2 and Code 2 Flash are available in GitHub Copilot; Code 2 Flash is also described as being in limited-access preview in Microsoft Foundry.
22
24 The available sources do not establish eligibility across every Copilot plan.
There is also a pricing caveat: reporting on Microsoft’s benchmark figures says the Code 2 cost comparisons used a promotional price.
16 The headline ratios therefore should not be treated as permanent list-price comparisons. And because the published percentages are relative cost claims, they are not a substitute for checking the applicable prices and usage terms for a specific deployment.
Microsoft’s broader AI strategy
Microsoft framed the launch as part of a broader effort to build its own AI capabilities alongside its competition with OpenAI and Anthropic. Axios reported that Microsoft AI chief Mustafa Suleyman said the company was making significant progress toward a general-purpose frontier model.
2 The materials available here do not give a release date for that model.
The clearest supported takeaway is narrower: Microsoft is using specialized models and lower-cost benchmark claims to make a price-performance case for MAI. Whether those models are the right fit depends on the task, the full pricing terms, and performance beyond the selected benchmarks.