A glowing central AI hub connects cloud servers to interfaces for coding, security, images, audio, healthcare, and workflows.
Microsoft used day one of Ignite 2026 to show that it wants to sell its own AI models as well as host everyone else's. Microsoft AI announced what it calls a family of seven new models, covering cybersecurity, coding, image generation, speech, healthcare transcription and decision routing. The pitch is not "we beat everyone." It is "we are close enough, for far less money."

Most of the numbers below are Microsoft's own claims. No independent validation has been published yet, and where the evidence is thinner, this article says so.

A glowing central AI hub connects cloud servers to interfaces for coding, security, images, audio, healthcare, and workflows. The "seven" models, and why the count is slippery​

Microsoft's list includes Dragon Copilot Transcribe, which is a product built on MAI Transcribe 2. It is not a standalone model launch. Maia 200 is an infrastructure announcement, not an eighth model. Axios counted six new AI models, which is probably the cleaner count if you only want new model families. The practical reading is seven items in the lineup, of which some are product integrations.

The lineup and Microsoft's claims​

ItemMicrosoft's headline claimAvailability noted in coverage
MAI Code 298% of GPT-6.1 Sol's performance at 40% of the price, and 96% of Claude Opus 5.5's at 5% of the costGitHub Copilot
MAI Code 2 FlashMatches GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20%GitHub Copilot; limited-access Foundry preview
MAI Cyber 2 FlashBeats GPT-6 Astra on a security benchmark at under 2% of the cost per taskReported as coming soon to Microsoft Security products
MAI Image 2.7#1 on Artificial Analysis image editing, at one-fifth the price of GPT Image 2.6 Sunburst (max)Public preview on Foundry, Mai Playground, OpenRouter, Genspark and Vercel
MAI Realtime 1Beats GPT-Realtime 2.1, from $0.01 per audio minuteInterest registration through MAI Labs
Dragon Copilot Transcribe (MAI Transcribe 2)10 million patient encounters a month, multilingual error rates halvedLive in Dragon Copilot
Decision 1.150x faster and 79x cheaper than GPT-Sol 6.1Described as generally available

Coding is the part most readers will touch first​

The most concrete news for developers is GitHub Copilot. GitHub's changelog says MAI Code 2 and MAI Code 2 Flash expand the MAI-Code family from lightweight assistance to complex, agentic software engineering. GitHub positions MAI Code 2 for deep debugging, multi-file implementation, refactoring and repository-wide reasoning. It positions Flash as lightweight and cost-efficient for everyday workflows.

Plan availability, per GitHub:

  • MAI Code 2: Copilot Pro, Pro+, Max, Business and Enterprise.
  • MAI Code 2 Flash: Copilot Free, Student, Pro, Pro+, Max, Business and Enterprise.
  • Surfaces: the changelog lists Copilot CLI, the GitHub Copilot cloud agent and the GitHub Copilot app.

Admin note: Copilot Enterprise and Business administrators can manage access to MAI models through the model policy in Copilot settings. In practice, the model may appear in the picker only after an admin enables the policy. If you run a Copilot tenant, check that setting before developers start asking why the picker looks different.

Rollout reporting differs by outlet. Constellation Research says MAI-Code-2 is generally available in GitHub Copilot. TokenPost said Microsoft gave no rollout date. GitHub's own changelog, dated November 17, reads as a live release, so GitHub is the better source. Staged rollouts can still differ by account.

What the benchmark numbers do and don't mean​

Microsoft says the coding figures are averages across four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1). It says these were chosen to represent everyday enterprise coding tasks. It also says MAI Code 2 is not yet tuned for harder work such as fundamental science research or very long-horizon agentic tasks. "98% of the quality" is therefore an average on a particular task mix, not a promise for every repository.

Price comparisons are also slippery. Microsoft's percentages depend on which list prices and token mixes it used, and the announcement doesn't fully spell that out. Treat them as a reason to run your own bake-off, not as a procurement spec.

TokenPost also reports that AT&T evaluations showed MAI Code 2 Flash outperforming certain GPT-6.1 variants at 50 times lower cost. It cautions that the comparison covered specific variants, not the whole GPT-6.1 family, and independent benchmarks have not verified the claims.

Security and healthcare​

MAI Cyber 2 Flash is aimed at monitoring and control agents that run continuously. Microsoft says it costs 86% less per output token than MAI Cyber 1.1 Flash. It also says MAI Cyber accounts for almost 40% of scans in its MDASH system. Microsoft's CyberGymE2E result is a vendor-reported benchmark. A model that does well on a benchmark and a model you trust to patch production are different things, so keep humans in the approval loop.

Dragon Copilot Transcribe is the most deployed item. Microsoft says it already handles transcription of doctor–patient conversations in the product, at 10 million encounters a month. Nothing in the announcement says the halved error rate applies to every language or setting, or that every clinician has access.

Decision 1.1, the unglamorous one​

Decision 1.1 is for routing, classification, prioritization, verification and workflow control. Axios-style headlines gravitate toward the flagship coding model, but this one may matter most for agent builders. If a small, cheap model decides when to call a big, expensive one, the cost savings come from architecture as much as from raw benchmarks. Microsoft reports 4,000 customers and says the model is generally available.

Maia 200 and MAI Licensed Weights​

Microsoft says Maia 200, its second-generation AI accelerator, is now in production. It says the chip is running its largest models with 1.7 times more tokens per watt. The comparison baseline isn't stated.

MAI Licensed Weights lets organizations customize MAI models and run them in their own Azure environments. Live coverage of the keynote described access for trusted partners through the Frontier program. The terms, eligible models and regions have not been published. Don't assume this is open-weights in the usual sense.

Why this matters​

  • Strategic: Axios notes that Microsoft has long relied on OpenAI and other AI companies for its most advanced models. Its own cheaper models give it leverage on both cost and supply.
  • Asked how much Copilot traffic would shift to MAI models, Mustafa Suleyman declined to give a target but predicted significant adoption.
  • Track record: MAI Code 1.1 Flash reportedly had 3 million users in GitHub Copilot, per Nadella's keynote remarks as relayed in live coverage.

What IT and dev teams should do now​

  1. Check your Copilot model policy settings to see whether MAI models are enabled for your org.
  2. Run MAI Code 2 Flash against your own routine tasks (explanations, targeted edits, small fixes) and compare cost and quality with your current default.
  3. Keep MAI Code 2 away from your hardest long-horizon agent work until you've tested it, given Microsoft's own caveat.
  4. For security tooling, wait for the Microsoft Security product integration before planning around Cyber 2 Flash.
  5. Ask Microsoft account teams about Licensed Weights terms before committing any roadmap to them.

Bottom line​

Microsoft is competing on price-performance rather than on being undisputed best. It is shipping where its distribution is strongest: GitHub Copilot, Foundry, Dragon Copilot and its security stack. The claims are bold and mostly unverified outside Microsoft. If they hold up, the cost pressure on OpenAI and Anthropic inside Microsoft's own products will be real.

 

References

  1. Microsoft takes on OpenAI and Anthropic with seven new AI models at Ignite 2026 - Neowin Neowin 2026-11-17T19:12:01+00:00
  2. MAI Code 2 and MAI Code 2 Flash in GitHub Copilot - GitHub Changelog github.blog
  3. Microsoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI models axios.com