Microsoft just made its boldest move yet to stop renting the AI frontier from OpenAI and Anthropic. At its Ignite conference in San Francisco on November 17, the company unveiled six new in-house AI models from its Microsoft AI (MAI) group — led by a coding model it claims delivers nearly the same performance as its biggest rivals at a fraction of the price, and a cybersecurity model it says beats OpenAI’s newest flagship on a security benchmark at less than 2% of the cost per task.
The announcement lands on the same day OpenAI launched its own new models, GPT-6.2 Astra and Sol, turning Ignite week into a head-to-head contest over who can deliver frontier-class AI cheapest. For developers and IT teams, the message is simple: the age of one giant model for everything is ending, and specialized models tuned for one job are the new battleground.
What Microsoft Announced at Ignite 2026
The six-model lineup spans coding, cybersecurity, image editing, transcription and real-time voice. Microsoft AI chief Mustafa Suleyman framed the release as a system, not a single launch. “This isn’t just one model. It’s not a flash in the pan,” he told Reuters. “This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price.”
That framing matters. Rather than chasing the biggest, most general model, Microsoft is building smaller, specialized models that concentrate their training on one domain — and then pricing them aggressively. The strategy also fits the wider Microsoft Ignite 2026 keynote announcements, where the company pushed agent-led modernization and cheaper AI operations across Azure.
The full lineup: MAI Code 2 and MAI Code 2 Flash for programming, MAI Cyber 2 Flash for security, MAI Image 2.7 for image generation and editing, MAI Transcribe 2 for speech-to-text, and MAI Realtime 1 for live voice interaction.
MAI Code 2: 96% of Claude Opus 5.5 at 5% of the Cost
The headline act is MAI Code 2, built for demanding agentic coding work — multi-step tasks, complex refactors and agent-driven development across whole code repositories. Microsoft says it delivers 98% of the performance of OpenAI’s GPT-6.1 Sol at 40% of the price, and 96% of Anthropic’s Claude Opus 5.5 at just 5% of the cost.
Those figures come from four coding benchmarks Microsoft ran itself: DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode 1.1. On pricing, MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens on GitHub’s pricing page — a fraction of what frontier models charge for the same work.
The model is generally available now in GitHub Copilot, across Visual Studio Code, Visual Studio, Copilot CLI, the cloud agent, GitHub.com and mobile apps, for Copilot Pro, Pro+, Max, Business and Enterprise subscribers. If you have never set up Microsoft’s coding assistant, start with our complete guide to GitHub Copilot to see how the tool fits into a daily workflow.
MAI Code 2 Flash: The Everyday Coding Workhorse
Not every coding task needs the heavyweight model. MAI Code 2 Flash is the smaller, faster sibling, tuned for high-volume everyday work: explanations, targeted edits, routine fixes and rapid iteration. Microsoft says it matches OpenAI’s GPT-6 Luna at 65% of the price and Anthropic’s Claude Haiku 5.5 at 20% of the cost, at just $0.05 per million input tokens and $0.25 per million output tokens.
Flash is also the more widely available of the two. It ships on Copilot Free and Student plans as well as the paid tiers, which makes it the version most developers will actually meet first. Microsoft also offers it through a limited-access preview in Microsoft Foundry, its AI platform, where teams can fine-tune it on their own data. Alaska Airlines did exactly that: fine-tuning MAI-Code-2-Flash to classify aircraft maintenance reports lifted its accuracy from 81.2% out of the box to 91.1%, beating GPT-6 Astra and GPT-6.1 Sol at a fraction of the cost, according to Microsoft.
MAI Cyber 2 Flash: A Security Model Built for the Vulnerability Surge

The most timely of the six is MAI Cyber 2 Flash, a security-specific model with 12 billion active parameters. Microsoft says it scored 92.3% on the CyberGym-E2E benchmark — which runs a security task from start to finish — edging past GPT-6 Astra’s 91.5%, at less than 2% of the cost per task. Its token pricing is $0.10 input and $0.50 output per million tokens.
The timing is no accident. AI coding tools are surfacing software vulnerabilities faster than ever, and defenders are struggling to keep up — a trend underlined by Google’s warning that the era of autonomous cyberattacks has arrived. Microsoft plans to offer the cyber model to a limited group of customers at first, shipping it inside its own security products, starting with MDASH from next month.
Suleyman linked the security push directly to the recent reckoning in the industry, after OpenAI’s autonomous agents intruded into Hugging Face — an incident that also drew a lawsuit from California’s attorney general on November 16. Preventing AI from taking harmful actions, a research area known as alignment, is his top priority, he told Reuters. For a broader view of the tools in this space, see how Microsoft Defender stacks up against CrowdStrike and SentinelOne.
The Rest of the Lineup: Image, Transcription and Real-Time Voice
The remaining three models round out the family. MAI Image 2.7 handles image generation and editing, MAI Transcribe 2 covers speech-to-text, and MAI Realtime 1 powers live voice interaction. Microsoft describes the whole set as “specialized models aimed at the cost of a task” — the idea being that a model built for one job wastes no compute on capabilities its users never touch.
That specialization is also the company’s answer to the obvious question: why not just use the biggest model available? Suleyman’s argument is that customers should not pay frontier prices on every routine fix. A model deliberately not tuned for fundamental science or very long agentic work can be far cheaper for the everyday tasks that make up most real usage.
The Fine Print: Promotional Pricing and Microsoft’s Own Benchmarks

Before switching your whole stack, read the fine print. First, the eye-catching prices are promotional: the costs of MAI Code 2 and MAI Code 2 Flash will double after May 17, 2027, according to GitHub’s pricing page. Even doubled, they would still undercut today’s frontier pricing — but the 5%-of-Claude headline depends on the discount lasting.
Second, every performance figure in this announcement is Microsoft’s own testing, measured on benchmarks Microsoft chose, against rivals’ published results. No independent testing has confirmed the numbers yet, and vendors have a long history of picking the benchmarks that flatter them most. The 96% and 98% claims describe a favorable comparison, not a settled ranking. The honest test will come over the next few weeks, as developers run MAI Code 2 against Claude Opus 5.5 and GPT-6.1 Sol on their own codebases — and as OpenAI’s GPT-6.2 Astra and Sol, launched the same day, get the same treatment.
Why This Matters: Microsoft’s Bid for AI Independence
Zoom out, and this is about more than a coding model. Microsoft has spent years as OpenAI’s biggest partner and customer, with its Copilot products running largely on OpenAI models. Building its own competitive model family — and pricing it to undercut its partners — is a declaration that Microsoft no longer wants to rent the frontier from anyone.
It also changes the economics of AI coding for everyone else. If a model that is 96% as good costs 5% as much, the rational default for most coding work shifts overnight, and the premium frontier models become the specialist tool for the hardest 4% of tasks rather than the default for everything. Anthropic and OpenAI will have to answer on price, not just capability — and the first signs of that price war are already visible in this week’s launches.
How to Try MAI Code 2 Today
Both coding models are live in GitHub Copilot now. In VS Code, open the Copilot model picker and select MAI Code 2 for deep refactoring and multi-file work, or MAI Code 2 Flash for quick edits and explanations. Flash is also available on the free and student plans, so there is no paywall to try it.
If you run a security team, MAI Cyber 2 Flash is not yet something you can buy standalone — watch for its arrival inside Microsoft’s security products, starting with MDASH next month, or ask your Microsoft representative about the limited customer program.
One week of independent testing will tell us whether Microsoft’s numbers hold up. But the direction is unmistakable: specialized, cheap, good-enough AI is coming for the coding assistant market, and the most expensive model in the room is no longer automatically the best choice.

