Microsoft announced a coordinated family of first-party AI models at Ignite on November 17, 2026, expanding its own model portfolio across six areas of enterprise work. The release includes MAI Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Image 2.7 Flash, Realtime 1, Transcribe 2 and Microsoft-Decision-1.1. Microsoft groups the package under a seven-model announcement, with the image Flash variant adding another separately named entry. (Microsoft’s Ignite announcement)
Availability is where the practical details sit. Some models can be selected inside existing Microsoft products, others are previews, and two are tied to specific product deployments. Treating the entire announcement as a generally available model catalog would give developers the wrong starting point.
The lineup also broadens Microsoft’s first-party AI offering beyond its OpenAI integrations. Its pitch centers on specialized models that can handle recurring enterprise tasks at lower cost, supported by comparisons that remain Microsoft’s claims.
Every Model and Its Launch Status
Microsoft’s seven-model framing mixes model names, variants and a transcription product. Dragon Copilot Transcribe is powered by MAI Transcribe 2, while Canvas is a workspace powered by Image 2.7 Flash. Neither product name needs to be counted as another underlying model.
The access map is more useful than the headline count:
| Model | Main role | Announced access or deployment |
|---|---|---|
| MAI Code 2 | Complex enterprise coding, debugging and repository work | Available in GitHub Copilot |
| MAI Code 2 Flash | High-volume coding assistance and targeted changes | Available in GitHub Copilot; limited preview in Microsoft Foundry |
| MAI Cyber 2 Flash | Specialist security work | Scheduled product rollout for eligible MDASH customers, starting in mid-December |
| MAI Image 2.7 | Image generation and editing | Public preview through Microsoft Foundry and selected partners |
| MAI Image 2.7 Flash | Faster, lower-cost image generation and editing | Public preview through Foundry and selected partners; powers Canvas |
| MAI Realtime 1 | Speech-to-speech conversations | Announced with starting pricing; access tier is unconfirmed here |
| MAI Transcribe 2 | Multilingual transcription | Deployed through Dragon Copilot Transcribe |
| Microsoft-Decision-1.1 | Structured decisions and workflow control | Generally available |
These distinctions come from Microsoft’s coding, security, image, voice and Ignite announcements.
A model being available inside Copilot says something different from an independently accessible Foundry endpoint. A scheduled upgrade for eligible security customers, meanwhile, gives developers no basis to assume a public API is available.
MAI Code 2 Splits Coding Into Two Workloads
Microsoft positions MAI Code 2 for work that requires understanding a larger codebase: deep debugging, multi-file implementation, repository-wide reasoning, production preparation and UI development.
Code 2 Flash targets more repetitive interactions, including code explanations, targeted edits, routine fixes and rapid iteration. Both are available in GitHub Copilot, but only Flash has an explicitly announced Foundry limited preview. (Coding announcement)
Microsoft’s published comparisons separate the two models:
| Model | Microsoft’s claimed performance | Microsoft’s claimed relative cost |
|---|---|---|
| Code 2 vs. GPT-6.1 Sol | 98% of benchmark performance | 40% of cost |
| Code 2 vs. Claude Opus 5.5 | 96% of benchmark performance | 5% of cost |
| Code 2 Flash vs. GPT-6 Luna | Comparable performance | 65% of cost |
| Code 2 Flash vs. Claude Haiku 5.5 | Comparable performance | 20% of cost |
The Code 2 figures average results across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Flash’s comparisons cover three benchmarks aimed at routine coding tasks. These are Microsoft-reported evaluations, with costs tied to the comparison workloads. They have not been independently verified in this reporting. (Benchmark details)
The distinction matters: “98% of performance” is a relative benchmark result, not a claim that the model completes 98% of production coding tasks correctly.
Microsoft also identifies a useful boundary. The coding models are not yet tuned for fundamental scientific research or very long-horizon agentic work. I’d use that boundary to shape an evaluation: test routine edits and multi-file changes separately, then measure accepted changes, review effort and retries. A lower-cost model becomes less attractive if maintainers spend longer repairing its output.
For related coverage on checking generated changes, see Zeniteq’s Claude Code verification playbook article.
Cyber 2 Flash Arrives Through a Security Rollout
MAI Cyber 2 Flash is Microsoft’s specialist security model for work such as vulnerability discovery, remediation recommendations and threat investigation. Microsoft says eligible MDASH customers will begin receiving it as a replacement for the existing MAI profile model in mid-December 2026. That makes this a scheduled product deployment, not a blanket public model release. (Security announcement)
Microsoft claims Cyber 2 Flash matches GPT-6 Astra’s quality on CyberGym-E2E, a vulnerability-discovery benchmark, at less than 2% of the cost per task. It also reports stronger results than GPT-5.6 Cyber on CyberGym-E2E and WebExploitBench at the same reasoning budget.
“Cost per task” is the consequential unit here. It reflects the resources consumed by the benchmark workflow; it cannot be substituted directly for a token-price discount or an expected reduction in a security team’s total bill.
For customers, the immediate question is eligibility and rollout timing. The benchmark claim becomes useful after confirming that the model is available in the security workflow they actually use.
Image 2.7 Adds a Flash Variant and Canvas
MAI Image 2.7 and Image 2.7 Flash enter public preview through Microsoft Foundry and selected partners. Microsoft is also introducing Canvas, an interactive workspace powered by Flash for creating, editing and refining images. (Image announcement)
Microsoft quotes approximate pricing of:
- Image 2.7: $35 per 1,000 images.
- Image 2.7 Flash: $10 per 1,000 images.
The company says Image 2.7 ranks first on Artificial Analysis image editing and outperforms GPT Image 2.6 Sunburst at its maximum setting for one-fifth of the cost. Those ranking and price comparisons are reported by Microsoft; the quoted per-image prices are approximate and depend on its stated pricing assumptions.
Using those figures, 100,000 images would correspond to approximately $3,500 for Image 2.7 or $1,000 for Flash. That is straightforward arithmetic from Microsoft’s estimates, not a production quote.
My view is that the two-model split makes particular sense for iterative image work. Teams could evaluate Flash for drafts and repeated edits, then assess whether the larger model produces enough additional usable outputs to justify its higher price. The useful measure is cost per accepted image, including discarded attempts.
Frequently Asked Questions
3 questions
1How Many Models Did Microsoft Announce?
Microsoft calls it a seven-model release. Counting Image 2.7 Flash separately and naming MAI Transcribe 2 behind Dragon Copilot Transcribe produces eight named model entries.