Microsoft used its Ignite conference on November 17 to introduce a group of seven AI models and services from Microsoft AI, covering cybersecurity, software development, image editing, live voice, transcription and decision routing. The announcement is part of a wider effort to give enterprises more options for choosing a model based on the task and its cost. It also comes with an important distinction: this is a collection of specialized releases, not one general-purpose system that Microsoft says can replace every other model.

What is in the seven-item lineup

The list names MAI Cyber 2 Flash for security work, MAI Code 2 and Code 2 Flash for coding, MAI Image 2.7 for image creation and editing, MAI Realtime 1 for live speech, Dragon Copilot Transcribe built on MAI Transcribe 2, and Decision 1.1 for routing tasks between models. Those products do different jobs and should be compared within their own categories. Microsoft separately described Maia 200, its second-generation AI accelerator, as being in production; the chip is infrastructure news and is not an eighth language model in the seven-item list.

The two coding models illustrate the intended split. Microsoft positions Code 2 for larger engineering tasks such as debugging, multi-file changes and repository reasoning. Code 2 Flash is aimed at frequent, narrower requests such as code explanations, targeted edits and routine fixes. Microsoft says both are available in GitHub Copilot, while Code 2 Flash also has a limited-access preview in Microsoft Foundry. The company says its flagship coding model is not yet tuned for harder scientific research or very long-running agent tasks.

Read the benchmark claims in context

Microsoft says Code 2 achieved 98 percent of GPT-6.1 Sol's performance at about 40 percent of its price, and 96 percent of Claude Opus 5.5's performance at 5 percent of the price, averaged across four coding benchmarks. For Code 2 Flash, the company reports comparable performance to GPT-6 Luna at 65 percent of the price and to Claude Haiku 5.5 at 20 percent. Microsoft also says MAI Image 2.7 ranks first on Artificial Analysis image editing and that Decision 1.1 is faster and less expensive than a named competing model.

These are useful comparisons to examine, but they are Microsoft's reported results, not a universal price list or a guarantee that every customer's workload will cost those amounts. Performance depends on the tests, task mix, model settings and scoring method. Image editing, coding and real-time audio also use different benchmarks, so the rankings cannot be combined into a single league table for all AI systems.

Availability is as important as model names

Several announcements connect the models to products people already use: Code 2 is offered through GitHub Copilot, Realtime 1 can be tried in MAI Labs, and Microsoft says Dragon Copilot Transcribe is in use for clinical documentation. Access differs by product and organization, so an announcement that a model exists does not mean every user can call it directly. Some Foundry access is described as a limited preview, while other features arrive inside a Microsoft service with its own eligibility and account requirements.

For developers and businesses, the practical decision is therefore less about counting models and more about checking the exact task, the service where the model is available, the usage price, and the evaluation that resembles their own work. Microsoft's release gives buyers more specialized choices, while independent workload tests and real deployment costs will determine whether its efficiency claims hold for a particular team.

FILEZUNA CONTEXT

Why it matters

The most consequential part of Microsoft's announcement is its task-specific strategy: a focused coding or routing model may be cheaper for routine work than sending every request to a premium general model. The trade-off is fragmentation. Buyers must understand which model is available in which product, what data terms apply, and whether a published benchmark resembles their codebase or workflow. Microsoft provides several benchmark and pricing comparisons, but they remain company claims tied to selected tests. A fair comparison for a real team should include the same prompts, model access conditions, latency, retries and total production bill, rather than relying on a headline percentage alone.

SOURCE NOTES

Sources and further reading

This story is written in its own words. These links identify the reporting and announcements used to verify the details above.

Facts checked Nov 21, 2026.

FZEND OF STORY
More Technology stories