Microsoft's Seven New MAI Models: Frontier AI at a Fraction of the Price

Amir Farid
7 min read

Microsoft's Seven New MAI Models: Frontier AI at a Fraction of the Price
Microsoft announced seven new MAI models at Ignite 2026 on 17 November 2026: MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, MAI Transcribe 2 and MAI Decision 1.1 — task-specialised models for enterprise workloads, pitched on price as much as capability.
The idea is simple: stop paying frontier prices for everyday jobs. If Microsoft's numbers hold up, the cheapest way to run AI in your business may soon be a different model for each job. Every figure here is Microsoft's own claim — and the fine print matters.
The seven models at a glance
All seven were announced together in Microsoft's Ignite 2026 keynote post:
- MAI Code 2: everyday enterprise coding — shipping code, debugging, multi-file changes.
- MAI Code 2 Flash: faster, cheaper coding for high-volume, lower-complexity tasks.
- MAI Cyber 2 Flash: security specialist for finding vulnerabilities.
- MAI Image 2.7: image generation, with a small reported leaderboard lead.
- MAI Realtime 1: Microsoft's first live speech-to-speech model.
- MAI Transcribe 2: speech-to-text for calls, meetings and recordings.
- MAI Decision 1.1: small, fast model for yes-or-no decisions inside agent workflows.
The pattern is deliberate: smaller models that each do one thing well — the logic behind our small-business AI agents guide.
MAI Code 2 and Code 2 Flash: the coding pair

The coding pair is the headline act. Both are generally available now in GitHub Copilot (model picker and Copilot Fusion) and in Microsoft Foundry. Microsoft positions Code 2 as the workhorse for everyday enterprise coding: understanding a large codebase and changing several files consistently. Our GitHub Copilot cloud automations guide explains how that work is increasingly done by agents that keep going after you log off — and a cheaper coding model changes the economics of every run.
The price claims — and the fine print
Microsoft's claims, stated as Microsoft's claims: Code 2 delivers 98% of GPT-6.1 Sol's coding quality at about 40% of the cost, and 96% of Claude Opus 5.5's at 5%, averaged over four benchmarks Microsoft chose (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, FrontierCode v1.1). Code 2 Flash is said to match GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost. See our Claude Fable 5.2 coverage and Cognition SWE-2 Flash guide.
The promo-price trap
Now the fine print: the cost figures use a 50% promotional price, and Microsoft printed no list price for Code 2 — so "40% of the cost" is what you pay during the promotion, not necessarily later. Build your business case twice: once at the promo price, once at double it. Microsoft says the coding models are not yet tuned for fundamental science research or very long agentic tasks.
MAI Cyber 2 Flash: security at a fraction of the cost
Cyber 2 Flash is the most striking claim of the seven. Microsoft says it scores 92.3% on the CyberGym-E2E security benchmark, against 91.5% for a leading frontier model — slightly better, at under 2% of the cost per task. It is priced at $0.10 per million input tokens and $0.50 per million output tokens.
If that holds up, it changes who can afford continuous security testing. Vulnerability scanning is high-volume, repetitive work, and the bill scales with volume. A model slightly better and roughly fifty times cheaper per task turns a job most companies do occasionally into one they can run on every commit. Read our OpenAI GPT-6 Cyber coverage, our Gemini 4 Carbon guide and our NVIDIA NemoClaw guide.
MAI Realtime 1 and MAI Transcribe 2: voice gets cheap
MAI Realtime 1 is Microsoft's first live speech-to-speech model — it listens and speaks in real time. Microsoft announced a starting price of $0.01 per audio minute — a ten-minute customer call costs about ten cents of model time. At that price the model stops being the expensive part of the call, which changes the economics of an AI receptionist completely. Transcribe 2 sits beside it, turning calls, meetings and recordings into text. Our step-by-step voice agent guide walks through building that pipeline; our AI chatbot vs AI agent guide helps you decide whether you need one at all. One caution: a starting price is not a final price, and per-minute pricing covers only the model.
MAI Image 2.7 and MAI Decision 1.1: the quieter two
MAI Image 2.7 is Microsoft's new image model, with a reported 4 Elo-point lead on the image leaderboard as of 16 November. Four Elo points is a small margin, so treat this as "competitive with the best", not "clearly ahead".
MAI Decision 1.1 is the smallest of the seven, and arguably the most interesting for anyone building agents: a small, fast model that answers yes or no, routes a request or classifies a document. Every agent makes dozens of small judgement calls per task, and paying frontier prices for each one is wasteful. Our Gemini Decision 1 coverage and our Sarvam Prism guide show the same idea from other vendors. And remember our GDPx coverage — no model clears roughly 12% on long-horizon agent tasks, so a decision model is a component, not a solution.
The hardware and the platform underneath
Maia 300 and Azure Cobalt 200
Microsoft announced the Maia 300 chip — 40% more tokens per watt than leading GPU-based systems, Microsoft says — and Azure Cobalt 200 VMs are now generally available. The cost advantage should come from Microsoft's silicon and cloud, not just from smaller models. Also announced: Copilot Workspaces, long-running Autopilot, Copilot Code, Microsoft Perception security agents, GPT-6.2 and Meta's latest models in Foundry.
Licensed weights and the lock-in question
Microsoft also announced MAI Licensed Weights: trusted customers can access Microsoft AI model weights through Frontier Co. — a concession to businesses that want to run models on their infrastructure, echoing the self-hosting story in our Mistral Large 4 guide and our self-hosted Claude managed agents coverage. But note the qualifier: "trusted customers", not an open download. The cheapest, most integrated path will be the MAI models on Microsoft hardware — a reasonable trade if you are already a Microsoft shop. Our du National HyperAI coverage makes the same point for UAE businesses: price sovereignty and control alongside the token bill.
What your business should do this week
Do not rebuild your stack around a keynote — run a cheap, honest comparison instead.
Step 1: pick one real job per model
Choose one real coding task, one security test and one voice call from your own work, not a demo. Run each on the MAI model and on your current model, and record success, time and cost.
Step 2: price the job at the list price, not the promo price
Re-run your numbers at double the promotional cost for Code 2 — no list price was printed. If the case still works, it is a real saving; if it only works at the promo price, treat the model as a trial. Our UK custom software cost guide and Dubai hiring guide help you budget the engineering time around the model.
Step 3: keep your exit open
Build the pilot so the model is swappable: prompts and evaluations in your own repo. If Microsoft's specialists are as cheap as claimed, you will want to keep them — but as a choice, not a trap. If you need help, our custom software development team can scope it in a week.
Frequently asked questions
What are Microsoft's seven new MAI models?
Announced at Ignite 2026 on 17 November 2026: MAI Code 2 and Code 2 Flash for coding, MAI Cyber 2 Flash for security, MAI Image 2.7 for images, MAI Realtime 1 for live voice, MAI Transcribe 2 for transcription and MAI Decision 1.1 for fast decisions.
How cheap is MAI Code 2, really?
Microsoft claims 98% of GPT-6.1 Sol's coding quality at about 40% of the cost, and 96% of Claude Opus 5.5's at 5% — but those cost figures use a 50% promotional price, and no list price was printed. Price your case at double the promo cost too.
Should my business switch to MAI models now?
Run a pilot first: one real coding task, one security test and one voice call, on the MAI models and on your current models, comparing success, time and cost at the list price, not the promo price. Keep the model swappable.