What Is Microsoft MAI Code 2? The Seven-Model Family, Explained
Microsoft MAI Code 2 is Microsoft’s bet that good-enough AI at 5% of the price beats the best AI at full price. Launched on November 17, 2026 at the company’s Ignite conference in San Francisco, Code 2 is the flagship coding model in a family of seven new MAI models — and it is already generally available inside GitHub Copilot, the coding assistant millions of developers use daily. Microsoft says it delivers 98% of OpenAI’s GPT-6.1 Sol coding quality at 40% of the price, and 96% of Anthropic’s Claude Opus 5.5 at just 5% of the cost. This guide explains what Microsoft MAI Code 2 is, what the other six models do, what they actually cost, how to use them, and where the headline claims need a second look.
All benchmark figures in this guide are Microsoft’s own unless labelled independent. No outside evaluation of the new models had been published at the time of writing, and that matters more than usual here, because the entire launch is built on price-performance claims.

What is Microsoft MAI Code 2?
MAI Code 2 is a specialised coding model built by Microsoft AI — the division led by Mustafa Suleyman — for the everyday enterprise coding work the company sees in GitHub Copilot: taking code to production, deep debugging, changes across multiple files, and refactoring across a whole repository. It is not a general-purpose frontier model. Microsoft is explicit that Code 2 and its lighter sibling are “explicitly not yet tuned for harder problems, like fundamental science research and very long-horizon agentic work”. The model is tuned for the routine, high-volume work that fills a developer’s day, and priced to be left running.
It is the second generation of Microsoft’s in-house coding line. At its Build conference in June 2026, Microsoft launched its first seven MAI models, including MAI-Code-1-Flash; the Ignite family is the follow-up, with a new coding pair, a new security model, a new image model, Microsoft’s first speech-to-speech model, a transcription model and a decision model. The strategy is specialisation: instead of one giant model for everything, a set of smaller models each built for one job, each priced far below the frontier general models.
The seven MAI models at a glance
| Model | What it does | Microsoft’s headline claim | Where to find it |
|---|---|---|---|
| MAI Code 2 | Everyday enterprise coding | 98% of GPT-6.1 Sol at 40% of the price; 96% of Claude Opus 5.5 at 5% | GitHub Copilot (generally available) |
| MAI Code 2 Flash | Quick edits, routine fixes, code explanations | Comparable to GPT-6 Luna at 65%; Claude Haiku 5.5 at 20% | GitHub Copilot; Foundry (limited preview) |
| MAI Cyber 2 Flash | Finding and patching vulnerabilities 24/7 | GPT-6 Astra quality at <2% of the cost per task | MDASH, eligible enterprise customers (mid-Dec) |
| MAI Image 2.7 | Image generation and editing | #1 on Artificial Analysis image editing, at 1/5 the price | Microsoft Foundry |
| MAI Realtime 1 | Live speech-to-speech conversation | Outperforms GPT-Realtime 2.1; from $0.01 per audio minute | MAI Labs |
| Dragon Copilot Transcribe (MAI Transcribe 2) | Medical transcription | 10M patient encounters a month; multilingual errors halved | Dragon Copilot (live) |
| Decision 1.1 | Routing tasks to bigger or smaller models | 50x faster, 79x cheaper than GPT-6.1 Sol | Microsoft Foundry |
MAI Code 2: the benchmarks behind the 96% claim
Code 2’s headline figures are averages across four coding benchmarks Microsoft chose to mirror the everyday tasks it sees in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. Averaged across those four, Microsoft says Code 2 reaches 98% of GPT-6.1 Sol’s coding quality at about 40% of the price, and 96% of Claude Opus 5.5’s performance at 5% of the price. Code 2 Flash, measured across three of the same benchmarks, matches GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20%.
Microsoft also reports a human test: in a blind, side-by-side evaluation of frontend work, developers preferred Code 2’s dashboards and visualisations over GPT-6 Sol’s 64% of the time. The model has been tuned to generate dashboards and visualisations, which is a deliberate choice — frontend polish is the part of coding work a developer judges by eye.
Three qualifications belong next to those numbers. First, the benchmarks are Microsoft’s choices, and an average across four tests can hide a test where the model trails badly. Second, the prices are promotional: both coding models are costed at a 50% promotional price, the promotion runs to May 17, 2027, and Microsoft’s posts print no list price for either model — so the 5% figure compares a promotional price with a rival’s list price. Third, no independent benchmark of Code 2 existed at launch. The honest summary is that Microsoft’s own tests show a capable everyday coding model at a very low price, and the size of the gap to the frontier models is unverified.
MAI Code 2 price and how to access it
There is no published list price for MAI Code 2 yet — only the promotional pricing behind the comparison claims. What is concrete is where it runs. Code 2 and Code 2 Flash are generally available in GitHub Copilot today, in the model picker and through Copilot Fusion, on Copilot Pro, Pro+, Max, Business and Enterprise plans. Code 2 Flash is also available on Copilot Free and Student plans. Code 2 Flash is additionally in limited-access preview in Microsoft Foundry, the company’s platform for building AI applications; Code 2’s Foundry availability is described differently across Microsoft’s own materials, so check the Foundry catalogue before planning around it.
To try it, open GitHub Copilot in Visual Studio Code, Visual Studio, the Copilot CLI, the cloud agent, GitHub.com or the mobile app, and pick MAI Code 2 from the model picker. Both models use usage-based billing for eligible subscribers on annual plans, according to GitHub’s release changelog.

The other six models
MAI Cyber 2 Flash is the security model, built for continuous monitoring and patching of live production systems. On CyberGymE2E, a benchmark of real-world vulnerability discovery, Microsoft says it completed 92.3% of tasks against 91.5% for GPT-6 Astra, at less than 2% of the cost per task — and that figure, unlike the coding claims, was costed at list prices. Its output tokens are priced 86% lower than the previous MAI Cyber 1.1 Flash. MAI Cyber models already account for almost 40% of scans in MDASH, Microsoft’s security scanning service, and 90% of scans that started with MAI continued to use it. From mid-December, Cyber 2 Flash replaces the older MAI engine in the MAI profile of MDASH for eligible customers. It is aimed at verified defenders, not general users, and it builds on the same security lineage as GPT-6 Cyber — the model it is measured against.
MAI Image 2.7 is the image model, and its claim is the most independently checkable of the seven: Microsoft says it takes the top spot on Artificial Analysis’s image-editing leaderboard, scoring 1,215 points against GPT Image 2.6 Sunburst’s 1,211, at a fifth of the price. At text-to-image generation it ranks fourth, making Microsoft the second-best lab on that measure. The smaller MAI Image 2.7 Flash brings image generation for high-volume use at about $10 per 1,000 images. Both are available in Microsoft Foundry, and Image 2.7 can be tried in MAI Labs, a limited preview for personal, non-commercial use.
MAI Realtime 1 is Microsoft’s first live speech-to-speech model, built for natural, low-latency conversation. In Microsoft’s evaluations it outperformed GPT-Realtime 2.1 across external and Dynamics 365 benchmarks, and it starts at $0.01 per audio minute. You can try it in MAI Labs.
Dragon Copilot Transcribe, built on MAI Transcribe 2, is already live in Dragon Copilot, where Microsoft says it powers 100% of doctor–patient conversation transcriptions — 10 million encounters a month, with multilingual error rates halved, in the hands of a third of US physicians. It is the least speculative model in the family: it is already doing the job at scale.
Decision 1.1 is a routing model — it judges which tasks need a big model and which a small one will do, so the expensive models are only used when needed. Microsoft says it is 50 times faster and 79 times cheaper than GPT-6.1 Sol, leads Google, OpenAI and the startup Jev on Artificial Analysis’s Decision Index, and is already used by 4,000 customers. It costs $0.042 per million tokens on the Global Standard tier. Note the name: this is version 1.1 of a product line whose first version reached general availability on Foundry in October, built on a different base model — the November release is a faster, multimodal update, not a first release.
Behind all seven sits Maia 200, Microsoft’s second-generation AI accelerator, now in production running the company’s largest models at 1.7 times more tokens per watt. Microsoft also introduced MAI Licensed Weights: take a frontier MAI model, customise it with your own data, and run it in your own Azure environment, with access under your control. AT&T’s custom model for 5G network operations, built under the earlier Custom Models programme, outperformed GPT-6.1 Sol at 50 times cheaper cost, according to Microsoft.
MAI Code 2 vs the rivals: where it actually stands
The honest comparison table is shorter than Microsoft’s, because only some of the figures are independently verifiable.
| MAI Code 2 | Claude Opus 5.5 | GPT-6.1 Sol | |
|---|---|---|---|
| Everyday coding (Microsoft’s 4-benchmark average) | 96% of Opus 5.5; 98% of GPT-6.1 Sol | Baseline | Baseline |
| Cost (Microsoft’s claim) | 5% of Opus 5.5; 40% of GPT-6.1 Sol (promotional) | List price | List price |
| Independent benchmark | None published yet | Leads several independent coding tests | Strong independent scores |
| Long-horizon agent work | Explicitly not tuned for it | Built for it | Built for it |
| Fundamental science research | Explicitly not tuned for it | Capable | Capable |
| Where it runs | GitHub Copilot | Claude Code, API, Copilot | ChatGPT Work, Codex, API |
The comparison that matters most is the one Microsoft is making with its own launch timing. OpenAI released GPT-6.2 Astra and Sol on the same day, with Astra 20% cheaper than its predecessor; Claude Fable 5.2 and Gemini 4 Hydrogen arrived within the previous week. Every lab is now competing on the price of near-frontier work, not just on who has the smartest model. Microsoft’s version of that bet is the most aggressive: specialised models, embedded in tools developers already pay for, at single-digit percentages of frontier prices. For background on the agent workloads all of these models are chasing, see our guide to agentic AI.
Limits and open questions
- No independent verification. Every performance figure is Microsoft’s own, on tests Microsoft chose. The Image 2.7 leaderboard claim is the exception — Artificial Analysis publishes its own rankings — but the coding and security numbers have no outside check yet.
- Promotional pricing. The 5%-of-Claude figure rests on a 50% promotional price with no published list price, and the promotion ends May 17, 2027. The real price comparison arrives after that date.
- Not a frontier replacement. Microsoft says so itself: the coding models are not tuned for fundamental science or very long-horizon agent work. A team doing frontier research or marathon autonomous runs still needs the big general models.
- Availability is uneven. Code 2 is generally available in Copilot, but Foundry access for Code 2 is described inconsistently, Image 2.7’s MAI Labs preview is limited and personal-use only, and Cyber 2 Flash is restricted to verified defenders through MDASH.
- The strategy risk. Specialisation is Microsoft’s bet that most coding work does not need frontier intelligence. If developers find the models fall short on the hard 10% of tasks, they will keep reaching for Claude or GPT-6 — and the price advantage will not matter.
Who should use MAI Code 2?
Consider it if you write code in GitHub Copilot every day and your work is the routine enterprise kind — debugging, refactoring, multi-file changes, explanations and fixes. It costs you nothing extra to try: it is in the model picker on the plan you already have, and if Microsoft’s figures are even roughly right, it does that work at a small fraction of the cost of the frontier models.
Look elsewhere if you need the best coding agent available for hard problems: Claude Opus 5.5 and GPT-6.1 Sol both have stronger independent results, and Code 2 is explicitly not built for the hardest work. Skip it too if you need a model you can run on your own hardware — MAI models run in Microsoft’s cloud, though Licensed Weights offers a customised, Azure-hosted middle ground.
Frequently asked questions
What is Microsoft MAI Code 2?
MAI Code 2 is Microsoft’s new coding model, launched November 17, 2026 at Ignite. It is built for everyday enterprise coding work in GitHub Copilot, where it is generally available now. Microsoft says it delivers 98% of GPT-6.1 Sol’s coding quality at 40% of the price.
How much does MAI Code 2 cost?
Microsoft has not published a list price. Its comparison claims use a 50% promotional price, running to May 17, 2027. It is available in GitHub Copilot on Pro, Pro+, Max, Business and Enterprise plans, with usage-based billing.
Is MAI Code 2 better than Claude Opus 5.5?
Not overall. Microsoft says it reaches 96% of Opus 5.5’s performance on four everyday coding benchmarks, at 5% of the cost — but those are Microsoft’s own tests at a promotional price, and Opus 5.5 leads on independent coding tests. Code 2 is also explicitly not tuned for very long-horizon agent work or fundamental science research.
How do I use MAI Code 2?
Open GitHub Copilot in Visual Studio Code, Visual Studio, the Copilot CLI, the cloud agent, GitHub.com or the mobile app, and select MAI Code 2 from the model picker. Code 2 Flash is also available on Copilot Free and Student plans.
What are the seven MAI models?
MAI Code 2 and Code 2 Flash (coding), MAI Cyber 2 Flash (security), MAI Image 2.7 and 2.7 Flash (images), MAI Realtime 1 (live speech), Dragon Copilot Transcribe on MAI Transcribe 2 (medical transcription), and Decision 1.1 (model routing).
Is MAI Code 2 available in Microsoft Foundry?
Code 2 Flash is in limited-access preview in Foundry. Microsoft’s own materials describe Code 2’s Foundry availability inconsistently — check the Foundry catalogue before planning around it.