Microsoft used Ignite 2026 on November 17 to release seven models under its MAI brand in a single batch, and the framing was consistent across all of them: not general intelligence, but a specific job at a specific price.

What shipped

The batch includes MAI Code 2 and MAI Code 2 Flash for enterprise coding, Cyber 2 Flash for security work, Image 2.7 for image generation, and Realtime 1 for live speech. The rest fill narrower slots in the same pattern — small, specialised, and priced to be used in bulk rather than demoed.

The claim worth testing

Microsoft says MAI Code 2 reaches about 98% of GPT-6.1 Sol's quality on coding tasks at roughly 40% of the price. If that holds on real repositories, it is the most commercially significant number in the release, because coding is the workload enterprises actually pay for at volume.

Two caveats. "98% of quality" is almost certainly an average across a benchmark suite, and coding quality is not linear — the last two percent is usually where multi-file refactors and subtle concurrency bugs live. And a 60% discount on a model that fails one task in twenty more often can easily cost more once you count the engineer time spent catching it.

Why task-specific models are winning

A frontier model is trained to be acceptable at everything, and you pay for that generality on every token. A coding model can drop the multimodal branches, the long-form writing tuning and the safety behaviours that only matter in consumer chat, and put the savings into price and latency.

The pattern also solves a procurement problem. Enterprises that cannot get comfortable sending code to a general-purpose assistant will often accept a model scoped to code and hosted in their own tenant. Scoping is a security argument as much as a cost one.

The security model

Cyber 2 Flash is the one to watch for a different reason. Security work is the clearest case where a model must be fast, cheap and heavily restricted — you want it reading logs at volume and you do not want it improvising. A model built for that job, rather than a general model prompted into it, is easier to argue about in an audit.

What is missing

No independent evaluation yet, and no published details on context window or tool-calling behaviour for the Code models. Tool calling matters more than raw quality for agentic coding, and vendors rarely lead with it.

Pricing beyond the headline ratio is also unclear — whether the 40% applies to cached input, output, or both changes the real comparison substantially.

How to test it

Take your repository's ten most recent bug-fix commits and replay them as tasks. Score on whether the patch applies and whether tests pass, not on how the diff reads. Then measure cost per merged patch, including the human review time. That number, not the 98%, decides whether the switch pays.