HomeAIAI ModelsMicrosoft Unveils MAI Code 2, Cyber 2 Flash and Perception

Microsoft Unveils MAI Code 2, Cyber 2 Flash and Perception

Microsoft has used its Ignite conference in San Francisco to launch MAI Code 2, MAI Code 2 Flash and MAI Cyber 2 Flash, and to say that Perception, its agentic security system, is now generally available.

In a 17 November 2026 post on Microsoft AI, the company said it was releasing “a family of seven new world class models.” The figures attached to those models are Microsoft’s own. “MAI Cyber 2 Flash matches GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGymE2E benchmark,” the post said. “MAI Code 2 reaches 98% of the coding quality of GPT-6.1 Sol at about 40% of the price. And against Opus 5.5 it reaches 96% performance for just 5% of the price.” A later passage gives the same Code 2 comparison as an average “over four widely used benchmarks” — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On the MAI Code 2 product page, Microsoft said Code 2 Flash “offers comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost, across three coding benchmarks measuring day-to-day, high-volume routine coding tasks.”

Microsoft also set a limit on what it says Code 2 is for. The models, it said, were built “to prioritize the coding work our customers do every day.” “They’re explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work.” It added that, in a blind side-by-side evaluation of frontend work, “developers preferred MAI Code 2 over GPT-6 Sol 64% of the time.”

GitHub’s changelog, dated 17 November 2026, 18:01 GMT, said MAI Code 2 is available in Copilot for Pro, Pro+, Max, Business and Enterprise plans, and that MAI Code 2 Flash is also available to Copilot Free and Student subscribers. Both, it said, are “available to eligible GitHub Copilot subscribers on annual plans.” Microsoft AI said Code 2 can also be tried in a limited-access preview in Microsoft Foundry.

On the cyber side, Microsoft said MAI Cyber “accounts for almost 40% of scans in MDASH,” and that Cyber 2 Flash is “priced 86% lower on output tokens than MAI Cyber 1.1 Flash.” In an edited version of Satya Nadella’s Ignite keynote, published on the Official Microsoft Blog, the company said it is “integrating Cyber 2 with Perception.” “Perception is a complete agentic security system with teams of red, blue and green agents that run through playbooks to deliver security outcomes at machine speed, with you always in control.” “Perception is now generally available.”

The rest of the seven-model lineup includes MAI Image 2.7, which TSN has already covered in its preview ranking for image editing, as well as MAI Realtime 1, which Microsoft said “starts at just $0.01 per audio minute,” Dragon Copilot Transcribe, Decision 1.1, and Maia 200, which the company again described as “now in production.” The chip claim is a continuation of Microsoft’s earlier Maia 200 and Maia 300 announcement. MDASH, the defender environment in which Microsoft says MAI Cyber already has a large share of scans, is the same programme TSN covered when the effort yielded 33 Linux kernel CVEs. Foundry remains the route Microsoft has been opening for its own agent stack, including the Copilot agent harness.

Nadella’s remarks also said Microsoft is “very excited to bring GPT 6.2 to Foundry.” That is a same-day placement of OpenAI’s GPT-6.2 Astra and Sol launch, not part of the MAI lineup, and is not merged into this story.

Microsoft’s price-performance claims are its own scorecard

The availability news is straightforward. Code 2 and Code 2 Flash are in GitHub Copilot for the plan tiers GitHub has listed, Perception is generally available on Microsoft’s own account, and Cyber 2 is being wired into that system. The percentages that will travel with the announcement are a different kind of fact. The 98% and 40% comparison with GPT-6.1 Sol, the 96% and 5% comparison with Claude Opus 5.5, the CyberGymE2E claim of Astra-level quality at less than 2% of the cost, and the Code 2 Flash comparison with GPT-6 Luna and Haiku 5.5 are all Microsoft’s figures, on benchmarks Microsoft has named, at prices Microsoft has used. They are not a bake-off TSN has run, and they should not be read as OpenAI’s or Anthropic’s numbers.

The same caution applies to the 64% frontend preference result and to the MDASH share figures. Those are Microsoft’s evaluations of its own models in its own products. They sit beside a limit the company has been unusually direct about: Code 2 is not, on Microsoft’s wording, yet tuned for fundamental science or very long-horizon agent work. Perception’s general availability is also an availability claim. Nadella’s description of red, blue and green agent teams is a product pitch for a system that is now on sale, not a measured record of how those agents have performed in customer environments. The useful follow-up is whether later independent coding and security evaluations, and customer use in Copilot and MDASH, support the scorecard Microsoft published at Ignite.

Sources

Related stories

Microsoft’s MDASH Effort Yields 33 Linux Kernel CVEs is the closest recent TSN account of Microsoft’s defender models in use, and is the next read.

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

  1. Oregon AG Files $2.035 Billion Action Against OpenAI
    AI Policy & Safety

More in this category

Latest stories

  1. Oregon AG Files $2.035 Billion Action Against OpenAI
    AI Policy & Safety

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.