Microsoft Unveils Seven MAI Models: Code 2 Matches Claude at 5% of the Price
Microsoft unveiled seven MAI models at Ignite on Nov 17, spanning coding, security, images, voice and transcription; Code 2 and Code 2 Flash are now generally available in GitHub Copilot.

Latest
1 updates · newest first-
Story upgrade: this flash was expanded in place into an in-depth story
Further updates will appear here
What happened
Microsoft Ignite 2026 opened in San Francisco on Nov 17, with CEO Satya Nadella singling out two model lines, Code 2 and Cyber 2 Flash, in his keynote; Microsoft AI published the full list of seven models on its website the same day, calling them “seven new world class models” and pitching scalable price-performance.
The seven models split by task: two for coding, one for security, one for images, two for voice and transcription, plus one decision model. Microsoft AI CEO Mustafa Suleyman told Reuters the company’s strategy is to build technology specific to different domains, keeping costs low while matching rivals.
Microsoft also announced two supporting moves the same day: its second-generation accelerator, Maia 200, is in production running Microsoft’s largest models at 1.7x more tokens per watt; and MAI Licensed Weights lets enterprises customize a frontier MAI model with their own data and run it in their own Azure environment. AT&T has used a custom model for 5G network operations analytics, which Microsoft says outperformed GPT-6.1 Sol at 50x lower cost.
Rollout channels are set: both coding models entered GitHub Copilot the same day, with Code 2 Flash also in limited-access preview in Foundry; Cyber 2 Flash is scheduled to reach a limited group of customers from mid-December; Image 2.7 and Realtime 1 are in Foundry, and Transcribe 2 is live across Dragon Copilot.
What each side said
Suleyman told Reuters: “This isn’t just one model. It’s not a flash in the pan. This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price.” He added that the recent incident in which OpenAI agents attacked the Hugging Face platform led him to focus on critical safeguards beyond performance, and that alignment was his top priority.
In an interview with Axios, Suleyman said specialization concentrates more relevant knowledge and expertise into a model of a given size; Microsoft AI wrote on its website that customers want models priced low enough to run agents 24-7 doing real engineering work.
Independent coverage adds caveats: Axios noted that the comparisons rely on benchmarks Microsoft selected and that some of the cost advantage comes from promotional discounts, with no outside validation yet; Decrypt noted Microsoft ran three of the tests on its own evaluation platform and used Artificial Analysis data for the fourth. OpenAI released GPT-6.2 Astra and Sol the same day, so Microsoft’s comparison still targets the previous generation, GPT-6.1 Sol.
The seven models
Coding leads the lineup. Code 2 targets demanding agentic coding work such as multi-step tasks, complex refactors and agent-driven development; its average pass rate across four benchmarks is 63.2%, against 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5, at an average cost of $0.309 per task versus $0.761 and $6.306. Prices use a 50% promotional price; GitHub’s pricing page lists Code 2 at $0.50 per million input tokens and $2.50 per million output tokens, with the promotional price set to double after May 17, 2027. Code 2 Flash handles high-volume everyday coding at a 61.0% average pass rate across three benchmarks and $0.052 per task; it is comparable to GPT-6 Luna at 65% of the price and to Claude Haiku 5.5 at 20%. Microsoft says neither coding model is yet tuned for fundamental science research or very long agentic tasks.
Cyber 2 Flash handles security. It completes 92.3% of tasks on the CyberGym-E2E benchmark against 91.5% for GPT-6 Astra, at under 2% of the cost per task, with that comparison costed at list prices; it has a one-million-token context window. MAI Cyber already accounts for almost 40% of scans in MDASH, and 90% of scans that started with MAI continued to use it; output tokens are priced 86% lower than the previous Cyber 1.1 Flash.
Images and voice each get a flagship. Image 2.7 tops Artificial Analysis’s image-editing leaderboard, leading by 4 Elo points as of Nov 16; the full model costs about $35 per 1,000 images against about $211 for GPT Image 2.6 Sunburst at maximum settings, while the lighter 2.7 Flash costs about $10 per 1,000 images; text-to-image ranks fourth. Realtime 1 is Microsoft’s first speech-to-speech model, starting at $0.01 per audio minute, which Microsoft says outperformed GPT-Realtime 2.1 across external and Dynamics 365 benchmarks. Transcribe 2 powers all doctor–patient conversation transcription in Dragon Copilot, covering 10 million encounters a month, halving error rates in multilingual workflows, and reaching a third of US physicians.
The decision model is Microsoft-Decision-1.1, built on Gemma, generally available and now supporting image inputs; Microsoft says it leads Jev, Google and OpenAI on the AA Decision Index, runs 50x faster and 79x cheaper than GPT-Sol 6.1, and is used by 4,000 customers.
Key facts
- Seven models: MAI Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Realtime 1, Transcribe 2 and Decision 1.1; Image 2.7 Flash is a lighter variant in the same family
- Code 2 scores: Average pass rate of 63.2% across four benchmarks, against 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5; $0.309 per task versus $0.761 and $6.306; prices use a 50% promotional price that doubles after May 17, 2027
- Code 2 Flash: Average pass rate of 61.0% across three benchmarks at $0.052 per task; comparable to GPT-6 Luna at 65% of the price and to Claude Haiku 5.5 at 20%
- Cyber 2 Flash: CyberGym-E2E at 92.3% versus 91.5% for GPT-6 Astra; under 2% of the cost per task at list prices; one-million-token context; rolling out to a limited group from mid-December
- Image 2.7: First on Artificial Analysis image editing, leading by 4 Elo points; about $35 per 1,000 images against about $211 for GPT Image 2.6 Sunburst at maximum settings; 2.7 Flash at about $10 per 1,000 images
- Voice and transcription: Realtime 1 starts at $0.01 per audio minute; Transcribe 2 covers 10 million encounters a month, halving multilingual error rates and reaching a third of US physicians
- Decision 1.1: 50x faster and 79x cheaper than GPT-Sol 6.1; used by 4,000 customers
- Hardware and platform: Second-generation accelerator Maia 200 is in production at 1.7x more tokens per watt; MAI Licensed Weights lets enterprises customize models with their own data and run them in their own Azure environment; Meta’s Muse Wonder, Muse Spark 1.3 and Muse Ember are coming to Foundry
- Caveats: All comparisons are Microsoft’s own, on benchmarks it picked and tests it ran; coding prices use a 50% promotional price; OpenAI released GPT-6.2 the same day, so the comparison targets the previous generation, GPT-6.1 Sol
Why it matters
Microsoft is now pitching its own models against the flagship products of OpenAI and Anthropic across coding, security and images at once, with per-task costs as low as 2% of its rivals'.
CompaniesOpenAI·Anthropic·Microsoft
Grab todayOpenAI Codex for Students: $100 in Free CreditsHIMIAO Picks →Follow HIMIAO Horizon
No sign-up · Unsubscribe any timeNotifications
A ping on your phone or computer when the brief is out. No email needed.
Comments
Related
- Nov 19MAI Code 2 vs Claude Opus 5.5 and GPT-6.1 Sol: $0.50 Input vs $4, Promo Price Until May 2027
- Oct 8Google Launches Universal Work Agent Gemini Agent, Running Gemini and Anthropic's Claude
- Nov 15SemiAnalysis Names Nvidia's Seven Biggest GPU Customers; Six Build Their Own Chips
- Nov 11Anthropic Gives Claude an Office Suite: Meetings, Sheets and Whiteboards
- Nov 9Four AI Giants Back a Common Opt-Out Signal for AI Training
Today
Nov 19 Thursday- Brief
- GitHubGitHub Weekly Vol.6: 7 Projects Worth a Look This Week
Tested or read — weeklyAll →
- Quiz
- Call
Latest news
All →- Nov 18Study: Frontier Models Can Be Tricked Into Marking Civilian Targets
- Nov 18Nvidia Reaffirms Open-Source Strategy, Hugging Face Deal on Track
- Nov 17Klarna Launches AI Assistant That Shops, Books and Runs Errands
- Nov 18Palo Alto to Buy Applied Compute for $4.5 Billion in All-Stock Deal
- Nov 18NGINX2Root Exposed: Five 9.2-Severity Flaws Let Attackers Run Code on Nginx Without Logging In
- Nov 18Nvidia Stops Shipping Licensed Hopper Chips to China
™