Microsoft Ships Seven MAI Models at Ignite: Code 2 Lands in Copilot, Claiming 96% of Claude's Performance at 5% of the Price
On 17 November 2026 Microsoft unveiled seven home-grown MAI models at Ignite, with MAI Code 2 and Code 2 Flash landing in GitHub Copilot the same day. Microsoft claims Code 2 reaches 98% of GPT-6.1 Sol's coding quality and 96% of Claude Opus 5.5's, at 40% and 5% of the price — but those comparisons use a 50% promotional price with no list price published, and benchmarks Microsoft chose itself.

In short: Microsoft is not racing to build the biggest model — it is racing to build the one that is good enough and cheap
On 17 November 2026, Microsoft used its Ignite conference in San Francisco to release seven home-grown MAI models in one go, covering coding, cybersecurity, image editing, live speech, medical transcription, and model routing. Microsoft AI CEO Mustafa Suleyman put it bluntly in the keynote: the mission is to build the best models in the world, but what customers really want is a price point that lets them run agents 24 hours a day doing real engineering work without the bill exploding. That sets the tone for the whole launch: not a contest over whose model is biggest, but over who can do everyday work for less money.
Here is the whole lineup in one sweep. Two coding models: MAI Code 2 and MAI Code 2 Flash. One security model: MAI Cyber 2 Flash, built for defensive vulnerability discovery and patching. Two image models: MAI Image 2.7 and its lighter sibling, Image 2.7 Flash. One speech model: MAI Realtime 1, Microsoft's first live speech-to-speech model. One medical transcription model: Dragon Copilot Transcribe, built on MAI Transcribe 2. And one routing model: Decision 1.1, which sends each request to the right-sized model and only calls on the expensive large model when it is really needed. Alongside the models came two supporting pieces: MAI Licensed Weights, which lets enterprises take model weights into their own Azure environment to customize and run them; and Maia 200, Microsoft's second-generation AI accelerator, already running its largest models in production at 1.7 times more tokens per watt.
For people who write code, the main act is Code 2. Microsoft says it is built for everyday enterprise coding work: shipping code to production, deep debugging, changes across many files, and reasoning across an entire repository. The performance numbers are reported like this: averaged over four benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1 — 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 just 5% of the price. Microsoft adds a qualifier worth quoting: both models are explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work. In other words, it is aimed at the code you actually write every day, not at research breakthroughs.
Frontend work is where Code 2 got special attention. Microsoft says it was tuned to generate great dashboards and visualizations, and in a blind, side-by-side evaluation developers preferred MAI Code 2's frontend work over GPT-6 Sol's 64% of the time. That sounds precise, but note the caveat: it is Microsoft's own blind test, and the sample size and judging criteria were not published. The reliable way to judge is the old one: take a page you actually need, run it on both models, and see which one feels better. Code 2 Flash handles the high-volume 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 for day-to-day routine coding.
Both coding models are already inside GitHub Copilot. GitHub's changelog from the same day spells it out: MAI Code 2 is available on Copilot Pro, Pro+, Max, Business, and Enterprise; Code 2 Flash also covers Copilot Free and Student. Both are billed on usage, and only to eligible subscribers on annual plans. The model picker covers a wide surface: Visual Studio Code, Visual Studio, the Copilot CLI, the cloud agent, the Copilot app, github.com, GitHub Mobile on iOS and Android, JetBrains, Xcode, and Eclipse. Code 2 Flash is also in limited-access preview in Microsoft Foundry. One date is worth marking: GitHub announced that with the arrival of Code 2 Flash, the previous generation MAI-Code-1.1-Flash will be retired from all Copilot experiences on 18 December 2026 — if your workflow still points at it, you have a month to move.
Hold on before switching: half of the price is a promotion
The most eye-catching part of these numbers is also the part that needs the biggest discount. Microsoft's own blog post prints no price list for Code 2, and several independent reports point out that the cost comparisons are calculated at a 50% promotional price: The Edge's reading is that both coding models are costed at their 50% promotional prices and the posts print no list price for either one. Axios likewise notes that some pricing advantages reflect promotional discounts. In other words, "40% of the price" only holds while the promotion lasts; once it ends, the same comparison could become 80%, and "5% of the price" could become 10%. That would still be cheap, but not as dramatic as the headlines make it sound. Microsoft has not said how long the promotion runs or what the list price is — so until independent evaluations land, the honest way to read the numbers is "at the promotional price", with a mental note that the list price is unknown.
The benchmarks themselves have a measurement problem too. The 98% and 96% figures are averages over four benchmarks Microsoft chose, run at maximum reasoning effort. Those four benchmarks target the everyday tasks Microsoft sees in GitHub Copilot — exactly the side of the story Microsoft wants you to see. Change the benchmark set or the task type and the ranking could easily change. On the very same day, OpenAI's GPT-6.2 launch used the same playbook — the vendor's scoreboard and the independent scoreboard disagreed — and that has become the norm this month. Microsoft's numbers have not yet been independently reproduced. Waiting for public leaderboards such as Artificial Analysis and Terminal-Bench to report results before moving your main workflow is the steadier course.
One more detail deserves attention: Microsoft did not publish hard parameters such as Code 2's context window or rate limits in this launch. For developers who feed whole repositories into a model every day, how much context it can hold and how much code it can process in one go directly decides whether it can take your work at all. Those numbers live in the model card or the Copilot documentation, not in the launch-day announcement. Checking them on the model page before choosing is more useful than staring at the 98% figure.
The security and image pieces may matter more than the coding ones
Focusing only on coding would miss two other solid parts of this launch. The first is MAI Cyber 2 Flash. Microsoft says it matches GPT-6 Astra's quality at less than 2% of the cost per task on CyberGymE2E, a benchmark of real-world vulnerability discovery. According to The Edge's reading of Microsoft's post, Cyber 2 Flash completed 92.3% of tasks on CyberGymE2E against 91.5% for GPT-6 Astra; on a second security test, WebExploitBench, it scored 40.4% against 38.4% for GPT-5.6 Cyber. Its output tokens are priced 86% lower than the previous MAI Cyber 1.1 Flash. Inside Microsoft's own multi-agent vulnerability scanning system, MDASH, MAI Cyber already accounts for almost 40% of scans, and 90% of scans that started with MAI continued to use it. But note the limit: Cyber 2 Flash goes only into Microsoft's own security products, automatically replacing the existing model for eligible MDASH customers from mid-December — the average developer will not meet it inside Copilot.
The second is images and speech. MAI Image 2.7 takes the number one spot on Artificial Analysis's image-editing evaluation; Microsoft says it outperforms GPT Image 2.6 Sunburst at maximum settings at just one-fifth of the price. On text-to-image generation it ranks fourth, which Microsoft describes as making it the second-best lab. The lighter Image 2.7 Flash targets high-volume scenarios at about $10 per 1,000 images. MAI Realtime 1 is Microsoft's first live speech-to-speech model; Microsoft says it outperformed GPT-Realtime 2.1 in its own evaluations, starting at $0.01 per audio minute. Dragon Copilot Transcribe is already running in the real world: it powers 100% of doctor-patient conversation transcriptions in Dragon Copilot, 10 million encounters a month, halves error rates in multilingual workflows, and is in the hands of a third of US physicians. Of all the numbers in this launch, the transcription one is the hardest — it is not a benchmark score, it is real usage already running.
Decision 1.1 deserves a mention of its own. It is a routing model: it sends each request to the right model so small jobs do not burn large-model money. Microsoft says it leads Jev, Google, and OpenAI on the Artificial Analysis Decision Index, is used by 4,000 customers, and is 50 times faster and 79 times cheaper than GPT-Sol 6.1. The model is generally available. For people doing vibe coding, the routing idea itself may be more useful than the model: if your project calls on several models, routing by task before calling anything saves more than picking a single cheaper flagship.
For Copilot users, the change arrives tomorrow
This launch differs from most model launches in one big way: it is not "available in a few months" but inside Copilot's model picker on the same day. If you already pay for Copilot, you will see Code 2 and Code 2 Flash in the dropdown without doing anything. Enterprise administrators should pay extra attention, though: GitHub says that under the default model-enablement setting, new models are enabled automatically unless an administrator has turned off the global default or explicitly disabled the model. If you manage Copilot policy for a team, check the model policy first and decide whether to let everyone try freely or lock it down pending evaluation.
The enterprise story sits with MAI Licensed Weights. Microsoft's pitch is: take a frontier MAI model, customize it with your own data, then run it securely in your own Azure environment — you control who can access it, and your data and adaptations stay yours. AT&T has already done it: it built a custom MAI model for 5G network operations analytics, and Microsoft says it outperformed GPT-6.1 Sol at 50 times lower cost. This line is aimed at customers who do not want to pour their intelligence into someone else's model. For individual developers and small teams it is out of reach for now, but it explains why Microsoft keeps investing in its own models: the more of its own models sit inside Copilot, the less it depends on OpenAI and Anthropic.
One more thing to note is the retirement of MAI-Code-1.1-Flash. Just last month we covered it: Microsoft squeezed a 137-billion-parameter home-grown coding model onto a local machine, with zero inference charges but a need for 120GB or more of memory. That piece was about the on-device version, while Code 2 is the cloud flagship — they are not the same thing. But the 18 December retirement date is real: if you use MAI-Code-1.1-Flash inside Copilot, move your workflow to Code 2 Flash or another model within a month. Microsoft's model generations are turning over faster than you might expect, and the previous generation's lifespan is shorter than it looks.
Why now: Microsoft is playing a bigger game
To understand this launch, look at where Microsoft has stood on models for the past two years. The flagship models behind Copilot have long come from OpenAI, with Claude also holding a seat inside Copilot. Microsoft is OpenAI's biggest partner and, at the same time, constrained by that partner: price rises, quotas, and roadmaps are not in its own hands. Reuters' report on 17 November made the point plainly: Microsoft is now moving to challenge OpenAI and Anthropic directly rather than continuing to act mainly as their distributor. Suleyman told Reuters that building technology specialized for different domains helps keep costs low while matching — and in some cases topping — model benchmarks. Translated: Microsoft no longer believes one giant model can do everything; it is betting on division of labor.
Whether that bet pays off depends on two things. The first is cost structure: Microsoft says customers want to run agents around the clock doing real engineering work, and token spend is skyrocketing — flagship general-purpose models cannot sustain that usage at their prices. Specialist models are smaller and more targeted, so they cost much less to run. Maia 200 is part of the same story: Microsoft's own accelerator delivers 1.7 times more tokens per watt, so the model and the chip save money together. The second is the capability boundary: Microsoft admits openly that Code 2 is not yet tuned for fundamental science research or very long horizon agentic work. The division of labor has a boundary — everyday engineering goes to specialist models, the hardest multi-day work stays with flagships. That boundary is honestly drawn, but it also means the 98% figure only holds inside the circle Microsoft drew; tasks outside it were never part of the comparison.
Another signal worth noting: Microsoft did not just release models, it released a customization line. MAI Licensed Weights lets enterprises take the weights into their own Azure environment, and AT&T's example — a custom model for 5G network operations analytics that Microsoft says outperformed GPT-6.1 Sol at 50 times lower cost — shows what is really being sold here: not performance, but control. Data stays in, changes stay yours, access is yours to grant. That appeals strongly to finance, healthcare, and telecom, where data compliance is sensitive, and it explains why Microsoft keeps funding its own models: what it is selling is "your models, your weights, your control". For individual developers that line is out of reach for now, but it decides which models Microsoft will push inside Copilot over the next few years — the more of its own models it has, the stronger its hand at the negotiating table.
What to try first: a comparison you can finish in a week
If you want to try now instead of waiting for independent evaluations, here is a comparison you can finish in a week. On day one, open Code 2 Flash in Copilot Free or Student and do the three things you do most every day: explain a piece of code you did not write, fix a small bug whose cause you already know, and add tests to a function. Track three numbers: how many rounds of rework each change takes, whether the result can be merged as-is, and how long you spend checking it. From day two, use Code 2 for cross-file work: refactor a module, add a feature across three files, or ask it to read the whole repository and propose a change plan. Track the same three numbers. Then run the same tasks once more on the flagship model you use today — and compare not who scores higher, but whose changes you would merge without hesitation and who costs you less checking time.
This comparison beats benchmark scores because benchmarks measure a model on a standard exam, while what really costs money in your workflow is rework and checking. A model that needs two rounds of rework on every change will eat the money it saved on price in your time, even at one-fifth the cost. The other way around, if Flash delivers usable results on everyday small jobs the first time, it is your cheapest sensible default. After a week you will have your own numbers: tokens per task, rework rounds, checking time. Those decide whether to switch your default far better than any percentage on a launch slide.
Our judgment: treat it as one more cheap option, not a generational change
What this launch really changes is not any single benchmark score but the fact that Copilot now has a set of "good enough and cheap" default options. The old logic was that even everyday tasks had to be picked from OpenAI's or Anthropic's flagships; now Microsoft says everyday tasks can run on its own models, two to four percentage points behind on performance at 40% of the price — or a fraction of it. If that claim holds, the monthly savings for a developer who runs agents dozens of times a day are visible. Flash landing in the Free and Student plans matters too: free users previously had only lightweight models, and now they can try Code 2 Flash directly inside Copilot.
Three signals are worth watching next. The first is independent evaluation: public leaderboards such as Artificial Analysis and Terminal-Bench will report Code 2's independent scores over the coming weeks, and then we will see whether the 98% figure stands or shrinks. The second is the list price: promotions always end, and Microsoft will eventually have to publish Code 2's official price list — that day's pricing is the real bottom line of this price war. The third is actual usage inside Copilot: GitHub has not said how many users will switch from GPT and Claude to MAI models. Only Microsoft knows that number, but it decides whether this launch merely added an option or genuinely rewrote Copilot's model landscape. Until those three signals arrive, the steadiest posture is: try it, but do not switch in a hurry.
Two cautions come first, though. One: do not make purchasing decisions from a vendor's benchmark scores and promotional prices. The 98% and 96% figures are averages over benchmarks Microsoft chose and ran itself; the 40% and 5% figures are calculated at a 50% promotional price, with neither the list price nor the promotion's end date published. Wait a week or two for independent leaderboards and real user feedback to settle, then decide whether to move your main workflow. Two: do not be dazzled by the "seven models" number. Only Code 2 and Code 2 Flash are inside Copilot and usable today; Cyber 2 Flash goes only into Microsoft's own security products, the image and speech models live in Foundry, and the routing model targets enterprises. For individual developers, this launch amounts to two cheaper options in Copilot's model picker — try Flash for everyday fixes, keep Code 2 for big cross-file changes, and leave genuinely hard, long-horizon agent work on the flagship models you have already verified.
Comments (0)
Sources
Related articles
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.
On 18 November 2026, Google Research and co-authors posted BuggyMLE to arXiv: bugs injected into 222 working ML repositories, and 14 models slide from 95% on the easiest tier to 9% on the hardest. A 9-billion-parameter model trained easy-to-hard beats Claude Opus 5.5 on the 10-bug tier.