Latest

Artificial Intelligence

Microsoft launches seven MAI models and takes aim at Astra and Opus

Ignite 2026 yields seven MAI models: Code 2 at 96% of Opus 5.5 for 5% of the price and Cyber 2 Flash under 2% of Astra's cost. What lies behind it.

Translated from Spanish by an automated system. The Spanish original prevails. Read the Spanish original

Microsoft launches seven MAI models and takes aim at Astra and Opus
Microsoft launches seven MAI models and takes aim at Astra and Opus

The screen on the Ignite stage showed two numbers and a ratio on November 17. MAI Code 2 reaches 96% of Claude Opus 5.5's performance in enterprise coding and sells at 5% of its price, the two figures that appear in the announcement of the seven MAI models. With those numbers, Microsoft stops arguing over who has the smartest model and moves on to arguing over how much a finished task costs, which is the purchasing department's question.

Ignite is the annual conference for technical customers and this year it served as a showcase for a complete family. There are two coding models and five more pieces: a cybersecurity specialist, an image one, a speech-to-speech one, a transcription one inside Dragon Copilot and a router that decides which model handles each request. None claims to be the most capable in absolute terms and all seven carry the price next to the message.

Three of the most repeated figures are 96%, 2% and 79: they come from tests chosen by the vendor and, in one case, from a promotional price that has not yet been finalized. None of them survives a rate change without the math shifting.

Seven models and a single argument: cost per task

Mustafa Suleyman, head of Microsoft AI, took the stage with an invoice on his mind. Token spending has soared and, at frontier prices, the numbers don't add up. A token is the unit used to bill a model's usage, and from there comes the shift toward specializing models by task instead of scaling a single one until it serves everything.

The November 17 list covers specific gaps. Cyber 2 Flash targets defense, Code 2 and Code 2 Flash target enterprise coding, Image 2.7 targets image editing, Realtime 1 targets spoken conversation, Transcribe 2 is integrated into Dragon Copilot and Decision 1.1 decides which model handles each request. They are seven products with the same commercial logic behind them.

That logic has a price, and the company admits it in writing: the models are built for a company's day-to-day work, not for the hardest problems. They are not tuned for basic scientific research or for agent tasks that last for hours.

A specialized model wins at the task it was trained for and loses when the problem goes off script. Microsoft turns that trade-off into a sales argument and leaves that work to other firms, which fill the gap with bets such as Nemotron for Science and its scientific agents. For a team that codes every day the deal is reasonable; for one seeking a new result, there is nothing to buy here.

Code 2, four tests and a 50% discount

The 50% promotional price is the piece that underpins the 96% in the announcement. Microsoft averaged DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, all enterprise coding and at maximum reasoning effort, and MAI Code 2 publishes 98% of GPT-6.1 Sol's performance at 40% of the cost and 96% of Opus 5.5's performance at 5% of the cost.

The detail lies in how each test was measured. Three ran on MSBench, Microsoft's internal evaluation platform, with the GitHub Copilot command-line interface and no subagents, while the FrontierCode 1.1 results and costs come from Artificial Analysis. Code 2's cost is also shown with its 50% promotional price, a footnote the company itself includes.

That discount does not invalidate the comparison, but it moves the floor on which it is built. If the 50% disappears, the 5% against Opus loses its basis and the announcement's sentence needs another figure to hold up.

There is one more measurement that almost no one repeats. In a blind evaluation of interfaces, games and visualizations, 258 raters preferred Code 2's output to GPT-6 Sol's 64% of the time. It is a human preference on aesthetics and functionality, not a measure of code correctness, and that is how the methodology note presents it.

Suleyman's argument went beyond price. If the same model serves to push the scientific frontier, that effort is paid for on every task, even when fixing a routine error, he said in the transcript of the MAI keynote. It is a reasonable justification for not chasing the capability ceiling and, at the same time, the admission that Code 2 is not the model for a new problem.

The Flash version of the model is already working inside the house. Microsoft derived MAI Work Flash from it, which serves several Microsoft 365 and FabricIQ products: in Word it uses 42% fewer input tokens per call than GPT-6.1 Sol and in Teams it spends up to 28% less on instructions and tool calls.

Cyber 2 Flash and the 92 tasks that came out of the benchmark

The advantage is eight tenths of a point and the cost is more than fifty times lower. In CyberGym-E2E, a vulnerability discovery test under real conditions, MAI Cyber 2 Flash achieves a 92.3% success rate versus 91.5% for GPT-6 Astra with half the reasoning budget.

The prices explain the rest. The model is billed at $0.10 per million input tokens, $0.01 cached and $0.50 output, versus $10, $1 and $50 for GPT-6 Astra in Azure OpenAI's global mode. The context window is one million tokens, four times that of the model it replaces.

The rollout is not open. It arrives exclusively in Microsoft's security products such as Codename MDASH and will replace the current model for eligible customers starting in mid-December, as the model note states.

Prior adoption helps explain the rollout. MAI Cyber accounts for nearly 40% of scans in MDASH and 90% of scans that start with an MAI model keep using it, while Cyber 2 Flash costs 86% less in output tokens than MAI Cyber 1.1 Flash.

The part that does credit to the announcement is the part almost no one publishes. When reviewing the results, Microsoft found that 92 of the 905 CyberGym-E2E tasks relied on the same open source projects and the same vulnerability reports that appear in its training data, as the MAI Cyber 2 Flash announcement itself acknowledges. It excluded those 92 tasks for all models, rivals included, and reports the result over the remaining 813.

A nuance remains that the big figure hides. Cost per task was measured on the CyberGym-E2E offline evaluation and at list price, not on a customer's actual spending, and the company clarifies this in the footnote. Insight Enterprises, one of the customers cited, reports 20% more token efficiency than the GPT models it tested.

Image, voice and transcription, the rest of the catalog

MAI Image 2.7 arrives with first place in Artificial Analysis's image editing ranking: 1,215 Elo points from 5,478 votes, four Elo points ahead of GPT Image 2.6 Sunburst at its maximum setting and a fifth of the price. The price comes to about $35 per 1,000 images and about $10 for the Flash version.

In text-to-image generation, MAI Image 2.7 is fourth and the company settles for being the ranking's second lab. It is the pattern across the whole family: win in a narrow category and charge less than the generalist leader.

The family has been Bing Image Creator's default option since June, with 14.7% compound weekly growth that Microsoft attributes to its own service, and it appears in tools from Adobe, Figma Weave and Manus. Living inside a search engine and an editor is worth more than a medal in a ranking.

MAI Realtime 1 is the house's first speech-to-speech model and starts at $0.01 per minute of audio, with reasoning billed separately at $1 and $5 per million tokens. The announcement compares MAI Realtime 1 with GPT Realtime 2.1 on Big Bench Audio and Tau-Voice and clarifies a practical detail for anyone who already has something built: it is compatible with OpenAI's Realtime API.

That means migrating an existing voice application does not require rewriting it. If you have a phone assistant in production, the change costs a configuration rather than a project, although the bill only drops when audio volume is high.

MAI Transcribe 2 is not sold on its own: inside Dragon Copilot it covers 100% of doctor-patient transcriptions, with 10 million encounters a month, half the errors in multilingual workflows and a third of US physicians as users.

Decision 1.1, routing as a product

Decision 1.1 is not a frontier model and does not pretend to be. It is a router that decides when the big model is needed and when a cheap one will do, and Microsoft places it ahead of Jev, Google and OpenAI in Artificial Analysis's decision index. It is used by 4,000 customers and the company puts it at 50 times faster and 79 times cheaper than GPT-6.1 Sol.

That 79 does not measure one model's intelligence against another's. It measures the savings from not always calling the most expensive one, and that is why it describes better than any other figure where the business is heading: selling less compute for the same task.

Model What it promises What it is compared with Relative price
MAI Code 2 96-98% in enterprise coding Claude Opus 5.5 and GPT-6.1 Sol 5% and 40% of the price
MAI Code 2 Flash Similar quality on routine tasks Claude Haiku 5.5 and GPT-6 Luna 20% and 65% of the price
MAI Cyber 2 Flash 92.3% in CyberGym-E2E GPT-6 Astra (91.5%) under 2% per task
MAI Image 2.7 First in image editing GPT Image 2.6 Sunburst max a fifth of the price
MAI Realtime 1 Low-latency speech-to-speech GPT Realtime 2.1 from $0.01 per minute
Decision 1.1 Routing between models GPT-6.1 Sol 79 times cheaper

The fine print the announcement does write

No comparison is false and all are favorable, which are two different things. Microsoft chooses the tests, sets the reasoning effort, decides which configuration of the rival to use and calculates cost with list or promotional prices depending on the case. A buyer who redoes the math with their own invoice will find similar numbers, not identical ones.

The gap between 96% and 100% is where complaints usually surface. Four points on an average of four tests means that on unusual tasks the cheap model fails more often, and those are precisely the ones that consume an engineer's time.

The competitive context does not disappear because the announcement does not mention it. Anthropic is preparing its IPO with the Claude Fable 5.2 and Mythos 5.2 models on the market. Google has spent months pushing Gemini 4 Hydrogen with flat pricing that attacks the same argument. The war is no longer fought over a ranking score, but over the monthly bill.

What Microsoft is gambling on with cost per task

The company has chosen ground where it can win without being the best. If the buyer compares cost per completed report or per hour of monitoring, the specialized model competes; if they compare raw capability on a problem no one has seen before, it loses. The November 17 announcement bets that the first kind of buyer is the majority at a mid-sized company.

One unknown remains that press releases do not resolve: whether Code 2's 50% promotional price becomes a stable rate or returns to its level at the end of the quarter. If you are buying coding capacity this quarter, that is the number worth watching, because when it changes, the comparison with Opus 5.5 will no longer be 5% and the math will have to be redone with the new prices.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
Hyundai Card and JPMorgan test currency exchange with AI agents and blockchain

Keep reading