One keynote lineup, several different states of ready

Count the launch the way the keynote counts it, per Thurrott's report, and the number is seven. Count it the way a team trying to use the models would, using Microsoft's own written posts, and the picture splits apart. Two coding models are live in GitHub Copilot today. A decision model is generally available — and is an update to a model more than 4,000 customers have already used, not a debut. Two image models are in public preview. A security model is promised through a scanner harness, with no date in the written announcements. The keynote presented one story, a family. The written record presents several, each at a different stage.

The coding pair is the most concrete part of the launch. MAI Code 2 is built for demanding agentic coding: multi-step tasks, complex refactors, agent-driven development. MAI Code 2 Flash is tuned for high-volume everyday coding. Both are generally available today on GitHub Copilot, in the model picker and as part of Copilot Fusion. Foundry access needs a qualification stated immediately: Nadella's post says both models are also available in Foundry, while the Foundry post specifically describes access to Code 2 Flash in Foundry as a limited-access preview. The two descriptions can coexist, but neither establishes generally available Foundry access for both models. Teams planning a Foundry deployment should check which statement their workspace reflects before assuming access.

The numbers attached to Code 2 are the kind that make a launch memorable, and they are also the kind that need a source attached. Thurrott reports from the keynote that Code 2 delivers 98% of GPT-6.1 Sol's coding quality at about 40% of the price, and 96% of Opus 5.5's performance for 5% of the price, and that Code 2 Flash matches GPT-6 Luna at 65% of the price and Haiku 5.5 at 20%. None of those figures appears in the two Microsoft written posts read for this piece. That does not make them false. It makes them unverifiable from the documents Microsoft published alongside the announcement, which is where a buyer would normally look for the terms they are being asked to trust.

The security and image claims follow the same pattern

Cyber 2 Flash is the sharpest example. Microsoft's keynote post calls it its most capable security model yet, one that delivers performance ahead of frontier models at a fraction of the cost per task. Thurrott reports the keynote figure: Cyber 2 Flash matches GPT-6 Astra at under 2% of the cost per task on the CyberGymE2E benchmark. The Foundry post describes the model in plainer terms: purpose-built for cybersecurity, spanning vulnerability discovery, validation and remediation, trained inside the Codename MDASH harness. It says the model will be available through MDASH. It gives no benchmark, no price and no date.

Image 2.7 follows the same shape. Thurrott reports it is now the top-rated model on Artificial Analysis for image editing, outperforming GPT Image 2.6 Sunburst (max) at one fifth of the price. The Foundry post says the image models bring sharper editing and more consistent generation, helping brands keep a visual identity from the first asset to the last, and that both Image 2.7 and Image 2.7 Flash are now in public preview in Foundry. The ranking claim and the price claim live in the keynote reporting. The preview status lives in the written post.

The remaining entries in Thurrott's lineup are the two audio models and the decision model. Thurrott reports that MAI Realtime 1 outperforms GPT-Realtime 2.1 and starts at $0.01 per audio minute; neither written post mentions Realtime 1, and its availability is not established in the sources read. Dragon Copilot Transcribe is a product built on MAI Transcribe 2; Thurrott reports it already powers 10 million patient encounters a month. Neither written post mentions the transcription product or its underlying model, so no access terms for either audio model can be taken from them. Decision 1.1 is the one model where the written record is fuller: generally available, built on Gemma, now supporting image inputs, used by more than 4,000 customers, and used by Microsoft itself for routing, data labeling, personalization and content filtering. Thurrott also reports Decision 1.1 as 50x faster and 79x cheaper than GPT-6.1 Sol; the reporting does not define the cost basis, and neither figure appears in the written posts.

The weights announcement is the one with the longest tail

MAI Licensed Weights may matter more than any single model in the set. Nadella's post puts it in one sentence: trusted customers will be able to get access to Microsoft AI model weights through the Frontier Co. The post states no deployment terms, no customization rights and no schedule. That is a notable direction for how Microsoft sells its own models — toward weights a customer can hold, not only API access — but everything beyond the sentence itself is unstated in the written record. The chip layer is similar. Per Thurrott's reporting, the models run in production on Maia 200. Maia 300 was announced the same day, and Microsoft projects it will deliver 40% more tokens per watt than leading GPU-based systems. That is a projection for a chip announced alongside the models, not a measured result.

There is also a video footnote worth noting for readers who follow this site for video models. Through a new partnership with fal on Foundry, Kling v3 Pro, MiniMax H3 Max, MiniMax H3 Max Turbo, LTX-2.5 Pro and LTX-2.5 Fast are coming to Foundry Models, per the Foundry post. That puts two of the video models this site tracks most closely inside Microsoft's catalog, alongside Meta's Muse models, which the Foundry post lists as coming soon.

The practical read is what buyers can establish from the evidence at hand. GitHub Copilot users can test Code 2 and Code 2 Flash now, on their own workloads, and that test is worth more than any keynote ratio. The reported price-performance figures need a test setup and a defined cost basis before anyone budgets around them; a figure published in a post would still need both. Cyber 2 Flash is a model to watch through MDASH, not one to budget around yet. Image 2.7 is a preview to evaluate, not a production switch.

Sources

  1. [1] Official Microsoft Blog (Satya Nadella keynote post), "Ignite 2026: Building the frontier ecosystem", November 17, 2026 (read first-hand in full by ALFW)Read source
  2. [2] Microsoft Tech Community (Foundry), "From model choice to your durable advantage", November 17, 2026 (read first-hand in full by ALFW)Read source
  3. [3] Thurrott, "Ignite 2026: Microsoft AI Announces 7 New Models", November 17, 2026 (keynote reporting: the seven-model lineup and all headline price-performance figures)Read source