Microsoft Unveils Seven New MAI Models at Ignite 2026 — Cheaper AI Takes on OpenAI and Anthropic

November 18, 2026

Microsoft Unveils Seven New MAI Models at Ignite 2026 — Cheaper AI Takes on OpenAI and Anthropic

Microsoft has made its biggest move yet to stop being only a seller of other companies’ artificial intelligence. At its Ignite 2026 conference in San Francisco on 17 November, the company unveiled seven new models built by its own Microsoft AI division, under the MAI name, covering coding, cybersecurity, image editing, voice, medical transcription and decision-making. The pitch was not that these models are the most powerful in the world. It was that they are good enough for everyday work at a much lower price.

That is a change of strategy, and it matters for the whole AI industry. For years Microsoft has been best known in AI as the main partner and distributor of OpenAI’s models, inside Copilot, Azure and GitHub Copilot. Now it is building a parallel family of its own, led by Microsoft AI chief Mustafa Suleyman, and selling it directly to businesses that run AI agents around the clock. The message to customers is simple: you do not need the most expensive model for every job, and you should not have to pay for one.

What Microsoft actually announced

The seven models are MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, Dragon Copilot Transcribe and Decision 1.1. Each one is aimed at a specific kind of work, rather than trying to do everything.

MAI Code 2 is the flagship. It is built for the coding that businesses do every day, including taking code into production, debugging across several files and making changes that span a whole project. Microsoft says it is already generally available inside GitHub Copilot, its coding assistant, from 17 November.

MAI Code 2 Flash is the smaller, faster version, built for high-volume, routine work such as explaining code, making small edits and fixing bugs. It is also in GitHub Copilot, with a limited preview in Microsoft Foundry, the company’s platform for building AI applications.

MAI Cyber 2 Flash is for security work. It is designed to find and patch vulnerabilities in live software, and to keep watch continuously rather than running a one-off scan. Microsoft says it will come only inside its own security products, starting with MDASH for eligible customers from mid-December.

MAI Image 2.7 is for creating and editing images. Microsoft says it ranks first on the Artificial Analysis image-editing leaderboard, and it can be tried in MAI Labs, a limited preview for personal, non-commercial use.

MAI Realtime 1 is Microsoft’s first speech-to-speech model, built for natural voice conversations with low delay, starting at $0.01 per minute of audio. Dragon Copilot Transcribe, built on MAI Transcribe 2, handles medical transcription and already supports 10 million patient visits a month, according to Microsoft, with errors in several languages cut in half.

Decision 1.1 is different from the rest. It does not write code or pictures. It decides which model should handle each task, sending simple jobs to a small, cheap model and only the hard ones to a large model. Microsoft says it is already used by 4,000 customers.

The price claims — and why they need care

The numbers Microsoft attached to the launch are bold, and they are the reason the announcement is being talked about. Microsoft says MAI Code 2 delivers 98 percent of the coding quality of OpenAI’s GPT-6.1 Sol at about 40 percent of the price, and 96 percent of Anthropic’s Claude Opus 5.5 performance at just 5 percent of the price. MAI Cyber 2 Flash, it says, matches GPT-6 Astra on a security test at less than 2 percent of the cost per task. Decision 1.1, it says, is 50 times faster and 79 times cheaper than GPT-6.1 Sol.

Those figures must be read with care, and Microsoft’s own reporting makes that clear. Every comparison is Microsoft’s own testing, measured on benchmarks Microsoft chose, and in the case of the coding models the price is a promotional price, not the full list price. Some of the savings also reflect discounts that run until 17 May 2027. That does not make the models bad, but it does mean the headline percentages are a favourable comparison, not an independent ranking. The models themselves, their names and where each one is available are confirmed facts. The savings are claims to be tested in real use.

That distinction matters for banks and other businesses in particular. A bank that runs AI agents to review documents, answer customers or check code all day pays for every token, so a model that is nearly as good at a fifth of the price could change its costs sharply. But a bank also cannot switch on a cheaper model for a regulated job on the strength of a vendor’s own benchmark alone. It will need to test the model on its own work, with its own data, before trusting it.

The chip and the control behind the price

Microsoft also announced that its second-generation AI accelerator, Maia 200, is now in production, running its largest models with 1.7 times more tokens per watt than leading graphics-processor infrastructure, according to the company. That is part of the cost story: cheaper models are easier to build when you also design the chip they run on.

It also announced MAI Licensed Weights, a programme that lets trusted customers take a frontier MAI model, tune it on their own data and run it in their own Azure environment. That is aimed at companies, including banks, that will not put their working knowledge into someone else’s model. Alongside the models, Microsoft announced GPT-6.2 coming to Microsoft Foundry, Meta’s latest models coming to Foundry, and a new product called Agent 365 Business Analytics.

The timing is also important. The launch came the same week as Nvidia’s strong earnings and Anthropic’s huge IPO pricing, with Anthropic expected to price on the evening of 18 November and start trading on Nasdaq as ANT on 19 November. Microsoft is not trying to beat those companies at their own game. It is trying to change the game: instead of one giant model for everything, a set of smaller, specialised models, each priced for the job it does.

What this means

For businesses, the Ignite announcement is a sign that the AI market is moving from a race for the biggest model to a race for the cheapest useful one. If Microsoft’s models perform in real work the way they perform on its own tests, the cost of running AI agents all day could fall sharply, and more companies could afford to use them. If they do not, the launch will still push OpenAI, Anthropic and Google to compete harder on price.

For Microsoft itself, the launch is a test of independence. It still sells its partners’ models, and it still works closely with OpenAI. But it now has a family of its own to offer when a customer asks for something cheaper, more specialised or easier to control. The next few months, as developers use MAI Code 2 in GitHub Copilot and businesses test the rest of the family, will show whether the price claims hold up outside Microsoft’s own laboratory.

This is not financial advice. Model performance and prices change, and readers should check the latest figures from Microsoft and independent testing before choosing a model for their own work.

Sources