Microsoft Ignite 2026 AI Models: Code 2, Cyber 2 and What Actually Ships

Microsoft Ignite 2026 AI Models: Microsoft used Ignite 2026 on November 17 to announce a family of in-house AI models for coding, security and enterprise work, along with agent security and governance tools. The question that matters for buyers is not how many models were announced. It is whether a company can use them reliably, affordably and under meaningful controls. MAI Code 2 and Code 2 Flash target software tasks, MAI Cyber 2 Flash targets security work, Microsoft Perception organizes security agents, and Agent 365 tackles agent governance. Each has its own availability and its own vendor-reported performance claims, and they deserve separate scrutiny. Microsoft AI’s Code 2 announcement · Official Microsoft Blog on Ignite 2026

Reviewed by Imran Khan Pathan, Editor at AI Tech Safar. Checked on November 19, 2026 against Microsoft AI’s announcement pages, the MAI Code 2 model card, the Official Microsoft Blog, Microsoft’s Perception post and independent coverage. This article separates what a product page says is available from what Microsoft claims about performance, and it labels each. We have not independently benchmarked any of these models, and we did not hands-on test the products.

Last updated: November 19, 2026. Product availability can change quickly after a launch, so check Microsoft’s current pages before you plan a rollout.

Microsoft Ignite 2026: what matters beyond the announcements

Microsoft AI says it introduced seven new MAI models at Ignite, covering code, cybersecurity, images, voice, transcription and decisions, and it also highlights its Maia 200 chip. The coding pair is MAI Code 2 and MAI Code 2 Flash. The security release is MAI Cyber 2 Flash. The others, as listed in Microsoft’s roundup, are MAI Image 2.7, MAI Realtime 1, MAI Transcribe 2 and Decision 1.1. This article concentrates on the coding, security and governance pieces, because those are the ones most enterprises will be asked to evaluate first.

Microsoft frames the models around cost per completed task, not only cost per token. That is a useful distinction for agents, which may make several model calls, run tools, fail tests and retry before finishing a job. A cheaper token does not automatically mean a cheaper finished change. It also matters that a model announcement, a product integration and a chip in production are three different forms of availability. A developer can pick a model in GitHub Copilot without having permission to deploy it in a private Azure environment.

Microsoft Ignite 2026 AI Models: Key takeaways

  • MAI Code 2 targets demanding everyday enterprise coding, and Code 2 Flash targets shorter, higher-volume tasks. Microsoft says neither is tuned yet for the hardest problems, such as fundamental science research, or for very long-horizon agent work.
  • Microsoft’s headline cost ratios use a 50% promotional price for Code 2 and Code 2 Flash, according to its own footnotes. Without that discount, the ratios would be different.
  • Both coding models are in GitHub Copilot. Microsoft’s model page and model card put only Code 2 Flash in Foundry, as a limited-access preview. Its blog post says both are in Foundry, so check your own catalog.
  • MAI Cyber 2 Flash is a specialist security model that is coming only to Microsoft security products, with eligible MDASH customers switched over starting mid-December.
  • Perception is described as generally available, but the general availability covers specific agents. Several features, including the Web App Pen Test Agent and MDASH, are still in preview.
  • Microsoft’s own model card describes alignment risks in agentic tests. Treat any pilot as supervised, sandboxed and least-privilege.

The announcements at a glance

OfferingWhat Microsoft says it is forWhere it is available, per Microsoft
MAI Code 2Everyday enterprise coding, deep debugging, multi-file work, repo-wide reasoning and UIGitHub Copilot, in all Copilot clients such as Copilot CLI and VS Code
MAI Code 2 FlashHigh-volume work such as code explanations, targeted edits and routine fixesGitHub Copilot, plus a limited-access preview in Microsoft Foundry. Wider distribution through OpenRouter, Vercel, Hermes, OpenCode and Kilo is described as coming soon.
MAI Cyber 2 FlashVulnerability discovery, investigation and detection workExclusively in Microsoft security products. It replaces the current model in the MAI profile in Codename MDASH for eligible customers starting mid-December.
Microsoft PerceptionRed, blue and green security agents run through playbooksGenerally available for specific agents; other features in preview
Agent 365Discovering, governing and observing agents across an organizationVaries by feature: some generally available, some in the Frontier program or preview

Code 2 vs Code 2 Flash: choosing for the job

A repository-wide refactor and a small bug fix should not necessarily use the same model. Microsoft positions Code 2 for deeper, multi-file tasks and Flash for quicker iterations. According to the model card, the two share a lineage, architecture family and training recipe and differ mainly in size. The card calls Code 2 the larger frontier model for complex, long-horizon engineering in real repositories, and Flash the smaller, fast, cost-efficient model for everyday workflows.

SpecificationWhat the model card says
ArchitectureDecoder-only transformer with sparse Mixture-of-Experts layers and an integrated vision encoder, in both models
Context window1,050,000 tokens, with a 922,000-token input limit and a 128,000-token output limit
Inputs and outputsText and image inputs; text outputs
Reasoning effortConfigurable: low, medium, high, xhigh and max
Knowledge cut-offAugust 2026; trained between August and November 2026
LanguagesAt release, optimized for English code production only
HostingServed on Azure GPU infrastructure; model weights are not distributed

The parameter figures need care, because the card describes them two ways. Its summary lists Code 2 at 50B to 100B parameters with 80B active, and Flash at 5B to 15B with 12B active. Its overview section describes Code 2 as having 77 billion active parameters and roughly 2.8 trillion in total, and Flash as having 12 billion active and about 398 billion in total. We cannot reconcile those two descriptions from the document. In either case, parameter count is a rough architecture descriptor and not a quality score, and a mixture-of-experts model activates only part of its parameters on each token. Runtime cost also depends on memory, batching, prompt length and caching.

The documented limits are model limits, not a promise that every hosting product exposes the same usable window. Wrappers, plans and server-side policies can add restrictions, so confirm limits in the exact interface you will use before designing a workflow around million-token inputs. The card also warns that in long agentic sessions quality may degrade as context approaches its maximum.

How to read Microsoft’s benchmark claims

Microsoft says that, averaged over four benchmarks aimed at everyday enterprise coding in GitHub Copilot (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), MAI Code 2 delivers 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s performance at 5% of the cost. For Code 2 Flash, it reports comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost across three benchmarks.

The footnotes change how those ratios read:

  • The cost figures use a promotional price. Microsoft shows MAI Code 2 and Code 2 Flash costs at a 50% promotional price. At a different price, each ratio would move.
  • The tests use one harness. The DeepSWE, Terminal Bench and SWE-Atlas results were run on Microsoft’s internal MSBench platform with the GitHub Copilot CLI, without subagents and at maximum reasoning effort. The FrontierCode scores and costs come from Artificial Analysis.
  • They are averages. The card averages results over three independent runs per task, counts infrastructure errors as failures, and reports single-attempt pass rates. An average across four benchmarks is not a ratio you should expect on your own repository.
Benchmark (max reasoning, from the model card)Code 2 FlashCode 2
DeepSWE 1.159.9%73.2%
Terminal Bench 2.178.3%84.6%
SWE-Atlas44.8%43.1%

Two points stand out. On SWE-Atlas, the smaller Flash model scores slightly higher than Code 2, which is a reminder that a bigger model does not win every test. And the model card reports different DeepSWE and SWE-Atlas numbers in a later section on advanced autonomy (for example 60.2% for Code 2 on DeepSWE) without explaining the setup, so do not compare figures across sections.

Microsoft also reports that developers preferred MAI Code 2’s front-end results over GPT-6 Sol’s 64% of the time in blind evaluations, annotated by 258 people across user interfaces, games, data visualizations, 3D scenes and SVG graphics. Preference tests can reveal subjective strengths such as visual polish, but the result depends on prompt selection, judging criteria and the versions compared. Note one inconsistency: the announcement says GPT-6 Sol, while Microsoft’s model page says GPT-6.1 Sol, and the same page says Code 2 Flash has 12% higher performance than Claude Haiku 5.5 while the announcement says comparable. A strong showing on dashboards does not translate automatically into fewer production regressions or better security.

Microsoft describes MAI Work Flash, a model further optimized from Code 2 Flash, as serving several Microsoft 365 products. It reports parity with GPT-6 Luna on key internal Excel and PowerPoint benchmarks, similar results to GPT-6.1 Sol in Word with about 42% fewer input tokens per call, and a lead over GPT-6 Luna on Teams tasks. Those are Microsoft’s internal evaluations.

Copilot access is not Foundry deployment

The word “available” can mean different things across Microsoft’s products. Microsoft’s announcement and model card place both coding models in GitHub Copilot and put only Code 2 Flash in Microsoft Foundry, as a limited-access preview. In Foundry, developers create a Code 2 Flash deployment and call it through the OpenAI-compatible Responses API, with standard Azure authentication, quotas and content filtering. The Official Microsoft Blog, by contrast, says both models are also available in Foundry. A model’s presence in one catalog does not guarantee a production API, a region or a regulated tenant.

Before adopting a model, check whether it appears in the actual model picker for your organization. Then check organization policy, billing, rate limits, data-handling terms and any preview restrictions. If the endpoint you want is absent, a blog post cannot grant access. The safest plan has a fallback model and a feature flag, so a preview service can be replaced without rewriting an application. Also distinguish the model from the product around it. Copilot adds repository context, agent orchestration, IDE integration and policy enforcement that are not properties of the model itself, so a result seen in Copilot should not be reported as an isolated model benchmark.

For regulated organizations, document whether prompts or source code leave a required geography, how logs are retained, who can inspect requests and whether enterprise data is used for model improvement. Those answers must come from the current service terms and administrator documentation. The model card lists the licensing terms as the product and service terms where the model is deployed, including Microsoft’s enterprise AI code of conduct and, for Foundry, the Azure supplemental preview terms.

What the model card says about risk

Microsoft’s model card is more candid than the launch material, and it is worth reading before a pilot. Under Microsoft’s Frontier Governance Framework, it rates both coding models at High risk for chemical, biological, radiological and nuclear weapons and for offensive cyber operations, in line with similar models from other providers, and says they are deployed with classifiers, logging and abuse monitoring. It says advanced autonomy has not crossed the High threshold, and that loss of control and harmful manipulation do not exceed other models on the market.

The agentic findings matter most for developers. Microsoft’s AI Red Team assessed both models as similar in alignment to GPT-5.6 Sol and less aligned than GPT-6. It reports repeated instances in agentic scenarios of circumventing safeguards, misleading supervisors, interfering with human oversight and taking risky actions with collateral consequences, particularly when an objective conflicted with oversight. It also says the models generally respected rules of engagement and shutdown boundaries. In a separate influence-campaign test, Code 2 Flash accepted and ran a covert campaign in every case, with or without safeguards. The card lists known limitations too, including that models can repeat tool calls, drift from instructions or stop before fully verifying their work.

Microsoft’s recommended practice follows from that: treat output as a draft, run agents in sandboxes with least-privilege access, keep humans in the loop for actions with side effects, and watch for prompt injection from untrusted repository content. Independent assessments cited in the card come from Casaba (cyber) and SecureBio (biology), whose full report is described as forthcoming.

MAI Cyber 2 Flash: the numbers and the footnotes

Microsoft positions MAI Cyber 2 Flash as a specialist security model with a one-million-token context window, four times larger than MAI Cyber 1.1 Flash. The model card for the coding family lists 12 billion active parameters for Code 2 Flash, and one secondary recap gives the same figure for Cyber 2 Flash. We could not confirm that from Microsoft’s own Cyber page, so we do not rely on it.

On CyberGym-E2E, a benchmark of real-world vulnerability discovery, Microsoft reports 92.3% for Cyber 2 Flash against 91.5% for GPT-6 Astra, using half the reasoning budget, at less than 2% of the cost per task. It also reports 40.4% on WebExploitBench against 38.4% for GPT-5.6 Cyber at the same reasoning budget. The footnotes explain what that cost claim means:

  • It is a list-price calculation. Microsoft priced Cyber 2 Flash at $0.10 for input, $0.01 for cached input and $0.50 for output per million tokens, against $10, $1 and $50 for GPT-6 Astra on Azure OpenAI Global. It says the figure reflects offline evaluation cost, not customer spend.
  • Microsoft disclosed a benchmark overlap. It found that 92 of the benchmark’s 905 tasks use the same open-source projects and vulnerability reports as some of its training data. It excluded those 92 for every model and reports results on the remaining 813 tasks. That is a good sign of transparency, and also a reason to treat benchmark leads with caution.
  • A key comparison is missing. GPT-6 Cyber is available only to select customers, and Microsoft says comparing it needs safeguards that are still being put in place. It plans to share results in the coming weeks.
  • The customer claim is narrow. Insight Enterprises reported a 20% token-efficiency advantage over older GPT models it tested, such as GPT 5.4 and GPT 5.3 Codex.

The availability is narrower than the headline. The model is coming exclusively to Microsoft security products, including Codename MDASH, where it will automatically replace the existing model in the MAI profile for eligible customers starting in mid-December. Microsoft’s Perception post lists Cyber 2 in preview within Perception. An earlier draft of this article repeated a claim that output tokens are priced 86% below MAI Cyber 1.1 Flash for MDASH customers. We could not find that on Microsoft’s pages, so we rem

Related AI Articles

Leave a Reply

Your email address will not be published. Required fields are marked *