On November 17, Microsoft presented a family of seven specialized AI models and systems at Ignite. The lineup includes MAI Code 2, Code 2 Flash, MAI Image 2.7 and the MAI Realtime 1 voice model. It also includes a security model, transcription in Dragon Copilot and Decision 1.1. The products have distinct access conditions and evaluation settings.

Matching a model to a job

Microsoft positions Code 2 for work across files and more involved changes, while Flash is intended for frequent small edits, explanations and repairs. The company publishes quality and cost comparisons against other models on selected benchmarks. These are vendor calculations that depend on tasks, cost definitions and configurations. Code 2 is listed in GitHub Copilot, and Code 2 Flash is also listed in limited-access preview in Microsoft Foundry.

For images, the MAI Image 2.7 card describes text-to-image generation and editing with an image input. This is a different workload. A team editing a catalog image, for instance, might care about whether everything outside one requested change stays intact. A benchmark for generating a whole image need not answer that question.

Comparing actual costs

For coding, measure accepted changes as well as request prices. If a lower-priced model requires more retries or more reviewer time, count that difference. Conversely, routine documentation or a small edit might not require the most expensive configuration at all.

Our editorial approach is to choose representative tasks from the team's own work and define acceptance criteria before running them. Compare cost, time, test outcomes and the human work required after the first answer. The announced specializations provide a starting point for that choice. Teams still need to check the actual product and terms of access.

In voice applications the number of minutes tells only part of the story. Value can depend on interruptions, comprehension accuracy and escalation to a human. Each modality needs its own acceptance criteria because a concise reply, a usable image and a correctly completed action are judged differently.

Editorial verification — 2026-11-18T10:21:35.465083+00:00