Microsoft launched seven of its own AI models at Ignite on November 17, 2026 — and the headline model, MAI Code 2, is already sitting in GitHub Copilot’s model picker today. Microsoft says Code 2 delivers 98% of OpenAI’s GPT-6.1 Sol coding quality at about 40% of the price, and 96% of Anthropic’s Claude Opus 5.5 at just 5% of the price.
That is a remarkable claim. It is also, entirely, Microsoft’s own claim — measured on benchmarks Microsoft picked, with the price side calculated at a 50% promotional discount. This article gives you the full picture: what each of the seven models does, what they actually cost, which Copilot plan gets you which model, and the two catches (the promo price and the vendor benchmarks) that decide whether switching is worth it for you.
The seven models, in one table
Microsoft AI, the division led by Mustafa Suleyman, announced the family as “seven new world class models” covering coding, security, images, voice, transcription and routing. Here is the lineup:
| Model | What it does | Where it lives | Key claim |
|---|---|---|---|
| MAI Code 2 | Everyday enterprise coding | GitHub Copilot (GA today) | 96% of Claude Opus 5.5 at 5% of the cost |
| MAI Code 2 Flash | Quick edits, routine fixes, explanations | GitHub Copilot (GA today), Foundry limited preview | Comparable to Claude Haiku 5.5 at 20% of the cost |
| MAI Cyber 2 Flash | Security agents, vulnerability discovery | Rolls out mid-December, verified defenders only | 92.3% on CyberGym-E2E vs 91.5% for GPT-6 Astra at under 2% of the cost per task |
| MAI Image 2.7 | Image generation and editing | Foundry public preview | #1 on Artificial Analysis image editing; ~$35 per 1,000 images |
| MAI Image 2.7 Flash | Cheaper image generation | Foundry public preview | ~$10 per 1,000 images |
| MAI Realtime 1 | Live speech-to-speech | Foundry | From $0.01 per audio minute |
| MAI Transcribe 2 | Speech transcription | Behind Dragon Copilot Transcribe | Already powering 10 million patient encounters a month |
Microsoft also announced Microsoft-Decision-1.1, a model-routing system now generally available, plus MAI Licensed Weights — a way for businesses to customize a frontier MAI model with their own data and run it inside their own Azure environment. And on the hardware side, Maia 300, its next AI chip, delivers 40% more tokens per watt than leading GPU systems, while Maia 200 is now in production.
The pattern is deliberate. Microsoft is not trying to win a “smartest model” contest. It is building a shelf of cheaper, specialized models for the work businesses actually do every day — and selling them inside products it already owns.
Code 2: the number that matters, and the asterisk under it
Code 2 is the model most readers will actually touch, because it is live in GitHub Copilot right now. Microsoft positions it for the everyday coding work that fills a developer’s week: taking code to production, deep debugging, changes across multiple files, reasoning across a whole repository, and building interfaces. It is also tuned for dashboards and visualizations — in blind frontend-app evaluations, reviewers preferred Code 2’s output over GPT-6 Sol’s 64% of the time.
The performance figures come from an average of four benchmarks Microsoft chose to mirror the tasks it sees in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average, Code 2 scores:
- 98% of GPT-6.1 Sol’s coding quality at about 40% of the price
- 96% of Claude Opus 5.5’s coding quality at 5% of the price
Microsoft published the per-task costs behind those percentages, and they are striking. At the promotional price, an average task costs $0.309 with Code 2, against $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5 — while the benchmark scores were 63.2% for Code 2, 64.1% for Sol and 65.3% for Opus 5.5. Read that carefully: Code 2 is not the best model on the chart. It is the model that gets within a couple of points of the best while costing a twentieth as much as Opus.
That is a genuinely useful trade for most coding work. The expensive model earns its keep on the hardest problems — a gnarly concurrency bug, a novel architecture decision, a security-critical refactor. The bulk of a working developer’s day is not those problems. It is writing a new endpoint, fixing a failing test, renaming a module across twelve files, or turning a sketch into a dashboard. If a model handles those at 96% quality for 5% of the price, the rational move is to use the cheap model for the bulk and keep the expensive one on standby.
The promo-price catch
Here is the asterisk, and it is a big one. Every price comparison Microsoft published for the coding models is calculated at 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, and Code 2 Flash at $0.05 input and $0.25 output — but those are the promotional rates. The prices double after May 17, 2027.
At full price, Code 2 would cost roughly four-fifths as much as GPT-6.1 Sol per task and about a tenth as much as Opus 5.5. That is still a strong deal — 96% of Opus at 10% of the price is nothing to shrug at — but it is not the “5%” headline. And Microsoft has not published what happens after May 2027: whether the promo simply ends, or whether list prices change again. If you are budgeting AI spend for the second half of 2027, plan on the doubled rate, not the launch rate.
The honest box, in one paragraph: all of these numbers are Microsoft’s own. Microsoft picked the benchmarks, ran three of the four on its own evaluation platform, and used Artificial Analysis results for the fourth. No independent testing has confirmed the results yet. Averages can also hide a weak spot — a model can average 96% while trailing badly on one specific kind of task. The right response is not to dismiss the launch, but to treat it the way you would treat any vendor benchmark: as a reason to test the model on your own code, not as proof.
Code 2 Flash: the cheap one you will actually use most
Code 2 Flash is the lighter sibling, built for high-volume routine work: code explanations, targeted edits, routine fixes and rapid iteration. Microsoft says it performs comparably to GPT-6 Luna at 65% of the price and to Claude Haiku 5.5 at 20% of the price, across three benchmarks measuring day-to-day coding tasks.
At $0.05 per million input tokens and $0.25 per million output tokens (promotional), Flash is priced for the kind of use where you stop thinking about cost entirely — the autocomplete-style, ask-a-quick-question, fix-this-typo layer of development. That is where most Copilot usage actually happens, and it is where a 5x or 20x price difference compounds fastest across a team.
Which Copilot plan gets which model
This is the part that decides whether you can try Code 2 today, and the answer depends on your plan:
| GitHub Copilot plan | MAI Code 2 | MAI Code 2 Flash |
|---|---|---|
| Free | No | Yes |
| Student | No | Yes |
| Pro ($10/mo) | Yes | Yes |
| Pro+ ($39/mo) | Yes | Yes |
| Max ($100/mo) | Yes | Yes |
| Business ($19/seat/mo) | Yes | Yes |
| Enterprise ($39/seat/mo) | Yes | Yes |
Both models are available in the Copilot model picker and through Copilot Fusion, across Visual Studio Code, Visual Studio, the Copilot CLI, JetBrains, Xcode, Eclipse, GitHub.com and the mobile apps, according to GitHub’s changelog. To switch, open the model picker in your editor’s Copilot chat panel and select MAI Code 2 or MAI Code 2 Flash — the same place you would pick Claude or GPT today.
One practical note: Copilot plans now bill in GitHub AI Credits (1 credit = $0.01) rather than premium requests, a change that took effect June 1, 2026. Code completions and next-edit suggestions stay unlimited and unbilled on paid plans; the meter runs on chat, agent mode, the CLI and cloud agents. If you want the full breakdown of how the credit system prices each plan, our guide to GitHub Copilot’s AI credits pricing walks through the real monthly bill at different usage levels.
The other five models, briefly
MAI Cyber 2 Flash is the most restricted of the seven. Microsoft says it succeeded on 92.3% of CyberGym-E2E’s vulnerability-discovery tasks, against 91.5% for GPT-6 Astra, at less than 2% of the cost per task. But you cannot just try it: access runs through Microsoft’s Codename MDASH program and only for verified defenders, with the rollout starting in mid-December. For most readers, this one is a watch item, not a tool.
MAI Image 2.7 and 2.7 Flash are in public preview on Microsoft Foundry, priced at about $35 and $10 per 1,000 images respectively. The comparison Microsoft wants you to notice: GPT Image 2.6 Sunburst at maximum settings costs roughly $211 per 1,000 images in Artificial Analysis’s comparison — about six times Image 2.7 and twenty times the Flash version. Image 2.7 also ranks #1 on Artificial Analysis’s image-editing leaderboard, though only by 4 Elo points as of November 16 — a lead, not a landslide. If you generate product mockups, marketing images or UI assets in volume, the price gap is the story; the quality gap is thin enough that your own test matters more than the leaderboard.
MAI Realtime 1 handles live speech-to-speech — the kind of model that powers a voice agent that listens and talks at the same time, interrupting and responding naturally instead of waiting its turn. It starts at $0.01 per audio minute. That is the model to watch if you are building phone agents, voice interfaces or live translation.
MAI Transcribe 2 is already working: it powers Dragon Copilot Transcribe, which Microsoft says handles 10 million patient encounters a month. Healthcare transcription is a niche for most readers, but it is the clearest proof in the launch that these models are not demos — at least one of them is already carrying production volume.
The bigger play: Microsoft is now a competitor to its own suppliers
There is a strategic story underneath the model list, and it matters for anyone choosing an AI stack. Microsoft is a partner and investor in both OpenAI and Anthropic, and it sells their models to its own customers through Azure. At Ignite, it stood on stage and pitched its own cheaper models against them.
Satya Nadella’s keynote made the positioning explicit: specialized models that concentrate the right knowledge into a smaller, cheaper package. Microsoft AI CEO Mustafa Suleyman told Axios the same thing in plainer terms — specialization is about getting particular work done economically, not about replacing every capability of a general-purpose model.
Microsoft also put a price on the whole Copilot stack. In the keynote, Copilot chief Jacob Andreou showed a chart of estimated monthly costs per business user, including subscription and usage fees: Copilot at $30 for an “everyday” AI user, against $81 for ChatGPT and $205 for Claude; for “advanced” users of each company’s agent tools, $96 for Copilot against $149 for ChatGPT and $299 for Claude. Those numbers come from research Microsoft commissioned and paid for from GigaOm, based on testing Microsoft ran in October on modeled workloads, and they reflect enterprise pricing with usage charges added to a per-seat fee — not the flat monthly subscriptions individuals pay. Treat them as Microsoft’s argument, not as an independent price survey. But even discounted for bias, the direction is clear: Microsoft wants the “good enough, much cheaper” lane, and it is pricing to own it.
The other strategic piece is MAI Licensed Weights. It lets a business take a frontier MAI model, customize it with its own data, and run it inside its own Azure environment. That is aimed squarely at companies that want model customization without sending their data to a third party — and it is the kind of offer OpenAI and Anthropic do not match in the same form. If your business has proprietary data and compliance requirements, Licensed Weights is worth a conversation with your Microsoft rep, even if Code 2 itself is not your priority.
Who should switch, and who should wait
Switch this week if: you write everyday enterprise code in GitHub Copilot and you are on a Pro, Pro+, Max, Business or Enterprise plan. Code 2 is already in your model picker. Run your normal work through it for a week — your real tasks, not a benchmark — and compare the results against Claude Opus 5.5 or GPT-6.1 Sol on the same tasks. The cost saving is real even at full price, and the only way to know whether the quality trade is acceptable is to measure it on your own codebase.
Try Flash if: you are on Copilot Free or Student, or you burn a lot of credits on quick questions and small edits. Flash is the cheapest way into the new family, and routine work is exactly where it is designed to shine.
Wait if: your work is fundamental science research, very long-horizon agentic tasks, or anything where a single wrong answer is expensive. Microsoft is explicit that neither coding model is tuned for those yet. Keep the frontier model for that work; use Code 2 for everything else.
Watch if: you are a security professional waiting on Cyber 2 Flash (mid-December, verified defenders only), or you generate images in volume and want to test Image 2.7 against GPT Image 2.6 on your own prompts before committing.
One more timing note: OpenAI also released GPT-6.2 on Tuesday, the same day as Ignite, which means Microsoft’s comparison against GPT-6.1 Sol is already one generation behind OpenAI’s newest model. That does not invalidate Code 2’s price case — the comparison is against a model that was current when the benchmarks ran — but if you are choosing between ecosystems rather than models, the GPT-6.2 numbers are the ones to watch next. Our pricing guide to the current flagship models, Gemini 4 Argon vs GPT-6 Astra vs Claude, covers how the top tier compares today.
What to do today
- Open GitHub Copilot and check the model picker. If you are on a paid plan, MAI Code 2 should be listed. If you are on Free or Student, Code 2 Flash should be there.
- Run a real task through it. Pick something from your actual backlog — a multi-file refactor, a failing test suite, a dashboard — and compare the output with what your usual model produces. Note the time and the credits used.
- Budget on the full price, not the promo. The 50% promotional price doubles after May 17, 2027. If Code 2 still wins at double the launch price — and on Microsoft’s own numbers it does, at roughly 10% of Opus’s per-task cost — then the switch is durable. If it only wins at the promo price, treat it as a trial.
- Do not move your hardest work yet. Keep frontier models for science, long-horizon agents and anything where a wrong answer costs real money. Code 2 is a workhorse, not a replacement for the whole stable.
- If you generate images at scale, test Image 2.7 in Foundry. At ~$35 per 1,000 images against ~$211 for GPT Image 2.6 Sunburst at max settings, a side-by-side test on your own prompts could pay for itself in a week.
Microsoft’s bet is that most AI work does not need the smartest model in the world — it needs a good model at a price that lets you use it without thinking. Code 2 is the first serious test of that bet inside the tool developers already use every day. The benchmarks say it is close. Your own code will tell you whether “close” is close enough.
Frequently Asked Questions
What are Microsoft’s MAI models?
MAI models are Microsoft’s own in-house AI models, built by Microsoft AI under Mustafa Suleyman. The seven launched at Ignite on November 17, 2026 are 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) and MAI Transcribe 2 (transcription). Microsoft also announced Microsoft-Decision-1.1 for model routing.
Is MAI Code 2 available in GitHub Copilot now?
Yes. MAI Code 2 and Code 2 Flash are generally available in GitHub Copilot as of November 17, 2026, in the model picker and through Copilot Fusion across VS Code, Visual Studio, the Copilot CLI, JetBrains, Xcode, Eclipse, GitHub.com and mobile apps. Code 2 is available on Pro, Pro+, Max, Business and Enterprise plans; Code 2 Flash is also available on Free and Student plans.
How much does MAI Code 2 cost?
GitHub’s pricing page lists Code 2 at $0.50 per million input tokens and $2.50 per million output tokens, and Code 2 Flash at $0.05 input and $0.25 output — but those are 50% promotional prices. The prices double after May 17, 2027. Microsoft says an average coding task costs $0.309 with Code 2 at the promo price, versus $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5.
Is Code 2 really as good as Claude Opus 5.5?
Microsoft says Code 2 delivers 96% of Claude Opus 5.5’s coding performance at 5% of the cost, averaged across four benchmarks Microsoft chose (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1). The benchmarks are Microsoft’s own and have not been independently verified, and the cost side uses the promotional price. On Microsoft’s own per-task numbers, Code 2 scored 63.2% against Opus’s 65.3% — close, not equal. Test it on your own code before deciding.
What is MAI Licensed Weights?
MAI Licensed Weights lets a business customize a frontier MAI model with its own data and run it inside its own Azure environment. It is aimed at companies that want model customization without sending proprietary data to a third-party provider. Microsoft announced it at Ignite on November 17, 2026, alongside the seven models.
Should I switch from Claude or GPT to MAI Code 2?
If you do everyday enterprise coding in GitHub Copilot, try Code 2 on your real work for a week and compare it with your current model on the same tasks. The cost saving is substantial even at the post-promo price. Keep a frontier model for fundamental science research, very long-horizon agent work, and anything where a single wrong answer is expensive — Microsoft says Code 2 is not yet tuned for those.
