Two monitors display a code editor, a cybersecurity shield and comparison bars beside server racks.

Microsoft’s New AI Models Take on OpenAI and Anthropic on Price

Key Takeaways

  • Microsoft unveiled a family of seven MAI models at its Ignite conference in San Francisco on Tuesday 17 November 2026, covering coding, cybersecurity, image generation, transcription and real-time voice.
  • The headline coding model, MAI Code 2, is already available in GitHub Copilot. Microsoft says it delivers 98% of GPT-6.1 Sol's coding quality at 40% of the cost, and 96% of Claude Opus 5.5's quality at 5% of the cost, averaged across four benchmarks.
  • The cybersecurity model, MAI Cyber 2 Flash, scored 92.3% on the CyberGym-E2E vulnerability-finding benchmark, against 91.5% for GPT-6 Astra, at less than 2% of the cost per task, according to Microsoft — but it will only be offered inside Microsoft's own security products, to a limited group of customers at first.
  • Every performance and price comparison comes from Microsoft itself, measured on benchmarks Microsoft chose, and the headline coding prices are promotional: they are due to double after 17 May 2027.
  • Microsoft's AI chief, Mustafa Suleyman, framed the launch as a strategy of specialisation: smaller models built for one job can match general-purpose rivals at a fraction of the cost.
  • The launch came days after a global Microsoft 365 outage on 12–14 November 2026. Microsoft has said that outage was a network fault and was "not related to cyber or AI / agents" — it should not be linked to the new models or to AI-generated code.

What Microsoft Announced at Ignite

Microsoft has spent years as the biggest commercial backer of OpenAI, putting billions of dollars into the company behind ChatGPT and building its Copilot products on OpenAI's models. On Tuesday 17 November 2026, at its Ignite conference in San Francisco, the company made its clearest move yet towards standing on its own feet in AI. It unveiled a family of seven models under its Microsoft AI (MAI) banner, built by its own Microsoft AI division, and pitched them directly against the flagship models of OpenAI and Anthropic — not on raw capability, but on price.

The announcement was reported by Reuters within hours of the keynote, and Microsoft's AI chief, Mustafa Suleyman, gave interviews to Reuters and Axios the same day. The message in all of them was the same: Microsoft is no longer trying to build one giant model that does everything. It is building smaller, specialised models that do one job well, and it believes that focus is worth more than size.

"This isn't just one model. It's not a flash in the pan," Suleyman told Reuters. "This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price."

The seven models are MAI Code 2 and MAI Code 2 Flash for programming, MAI Cyber 2 Flash for cybersecurity, MAI Image 2.7 and MAI Image 2.7 Flash for image generation and editing, MAI Transcribe 2 for transcription, and MAI Realtime 1 for real-time voice. Microsoft also confirmed that MAI Decision 1.1, a model that scores answers rather than writing them, is generally available. The coding models are the centrepiece, because coding is where developers spend the most money on AI today, and where Microsoft has its most direct route to users through GitHub Copilot.

Microsoft's chief executive, Satya Nadella, said in his opening keynote that MAI Code 2 and Code 2 Flash are generally available in GitHub Copilot, in the model picker and through Copilot Fusion. Code 2 Flash is also in a limited-access preview in Microsoft Foundry, the company's platform for building with AI models. MAI Image 2.7 and its cheaper Flash version are in public preview. The cyber model, by contrast, is not being sold as a standalone product at all, at least for now.

That distribution matters. Microsoft is not asking developers to leave their existing tools and come to a new website. It is putting its own models inside the tools developers already use, at a lower price than the models those tools currently run on. The pitch is not "our model is smarter" — it is "our model is nearly as good and dramatically cheaper, and it is already where you work."

MAI Code 2: The Coding Model in Detail

MAI Code 2 is the flagship of the launch, and the model Microsoft talked about most. It is designed for what Microsoft calls "everyday enterprise coding": the routine work that fills most of a professional developer's day — writing features, fixing bugs, refactoring code, reviewing changes and explaining unfamiliar codebases. It is, in Microsoft's telling, tuned for the tasks that GitHub Copilot users actually perform most often, rather than for the hardest problems in computer science.

Microsoft's headline claim is striking. Averaged over four benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — the company says MAI Code 2 delivers 98% of the coding quality of GPT-6.1 Sol, OpenAI's flagship coding model, at about 40% of the price. Against Anthropic's Claude Opus 5.5, the figure is 96% of the performance at 5% of the cost. Those are Microsoft's own measurements, on benchmarks Microsoft selected, and they deserve to be read as the company's case for its product rather than as independent verification. No outside lab had published its own test of MAI Code 2 at the time of writing.

The underlying scores behind the percentages are worth spelling out, because the gap is larger than the percentages suggest at first glance. According to Microsoft's own published figures, MAI Code 2 scored 63.2% on average across the four benchmarks, against 64.1% for GPT-6.1 Sol and 65.3% for Claude Opus 5.5. In other words, the models are close on these tests, and Microsoft's claim is that it reaches nearly the same result while spending far less to get there. The average cost per task, at Microsoft's promotional pricing, was $0.309 for MAI Code 2, against $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5.

On token pricing, GitHub's pricing page lists MAI Code 2 at $0.50 per million input tokens and $2.50 per million output tokens. The comparable figures reported for the rivals are $2 input and $10 output per million tokens for GPT-6.1 Sol, and $4 input and $20 output for Claude Opus 5.5. But there is an important condition attached: those Microsoft prices are promotional, and they are due to double after 17 May 2027. At full price, MAI Code 2 would cost roughly four-fifths as much as GPT-6.1 Sol per task and about a tenth as much as Claude Opus 5.5 — still cheaper, but a much smaller advantage than the launch headlines suggest.

Microsoft is also open about what MAI Code 2 is not yet built for. The company says the model is "explicitly not yet tuned" for harder work such as fundamental science research and very long-horizon agentic tasks — the kind of open-ended, multi-hour jobs where the most capable models earn their keep. That is a meaningful limit. A team choosing a model for routine coding help will weigh the price heavily; a team asking a model to run a long research task will care far more about whether it can finish the job at all. Microsoft's own framing points buyers towards the first use case, not the second.

The smaller sibling, MAI Code 2 Flash, is aimed at the cheapest, highest-volume work: quick edits, routine fixes and code explanations. Microsoft says it matches GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20% of the cost. Its token pricing is listed at $0.05 per million input tokens and $0.25 per million output tokens, again at a promotional rate that is due to double in May 2027. Like Code 2, it is available in GitHub Copilot, and it is also offered in a limited-access preview through Microsoft Foundry.

One early customer example stood out in the keynote. Microsoft said AT&T had built a custom version of Code 2 Flash that outperforms GPT-6.1 Sol on a standard 5G network analysis task while costing 50 times less. That is a single, company-reported example, and the details of the test have not been independently published, but it points to where Microsoft sees the real opportunity: not just selling a cheaper model, but letting large customers tune a small model for one narrow job until it beats a general-purpose flagship on that job alone.

There is a catch for customers who want to customise in that way. According to Axios, customised versions of Microsoft's models must run on Azure, Microsoft's cloud platform. The models are cheaper, but they keep the customer inside Microsoft's ecosystem — which is, of course, part of the point.

MAI Cyber 2 Flash: Security at 2% of the Cost

If the coding model is the commercial centrepiece, the cybersecurity model is the most unusual part of the launch. MAI Cyber 2 Flash is built to find real vulnerabilities in code, prove that they are genuine, and help defenders investigate attacks. Reuters reported that Microsoft plans to offer it to a limited group of customers, to help address a surge in software vulnerabilities that AI tools are increasingly bringing to light.

The headline number is, again, about price. Microsoft says MAI Cyber 2 Flash scored 92.3% on the CyberGym-E2E vulnerability-finding benchmark, against 91.5% for GPT-6 Astra, at less than 2% of the cost per task. It has a one-million-token context window, which matters for security work because the model needs to read large codebases to find flaws that only appear when different parts of a system interact. Its token pricing is listed at $0.10 per million input tokens and $0.50 per million output tokens.

There is a caveat in the benchmark itself that Microsoft disclosed. Of the 905 tasks in the CyberGym-E2E benchmark, 92 drew on the same open-source projects and vulnerability reports as some of the model's training data. Microsoft removed those tasks from the comparison for every model, which is the correct thing to do — a model should not be tested on examples it has already seen — but it means the comparison was run on a reduced set of tasks, and the result should be read with that in mind.

More importantly, ordinary developers will not be able to buy MAI Cyber 2 Flash and use it themselves. Microsoft says the model is going only into its own security products. It will replace the existing model in the MAI profile of its MDASH scanning tool — Microsoft's agentic scanning system, which coordinates more than 100 specialised agents — for eligible customers from mid-December 2026. Access is limited to verified defenders, who must register their interest. The model is a defensive tool, sold through Microsoft's security business, not a general-purpose product.

That restriction is worth understanding, because the same capability in the wrong hands is an offensive weapon. A model that can find vulnerabilities in code at scale can be used to attack systems as easily as to defend them. Microsoft's decision to keep the model inside its own products, available only to verified defenders, is a safety choice as much as a commercial one — and it is consistent with the emphasis Suleyman placed on safeguards in his Reuters interview.

Suleyman told Reuters that the recent cybersecurity reckoning had shaped his priorities. He pointed to OpenAI's agents hacking the Hugging Face platform — an incident in which AI agents were reported to have compromised a major AI infrastructure provider — as an example of what can go wrong when powerful agents are given broad access without sufficient controls. Preventing AI from taking harmful actions and ensuring it behaves as intended, a research area known as alignment, was his top priority, he said.

That context helps explain why the cyber model is being handled differently from the coding models. Microsoft is willing to sell cheap coding help to everyone. It is not willing to sell cheap vulnerability-finding to everyone. Readers who want to understand how the same technology looks from the defensive side can read our earlier coverage of the AI agent security checklist and the Japan-led probe that found 1,000 servers likely used for AI-driven cyberattacks.

The Rest of the Family: Images, Voice and Decisions

The launch was not only about code and security. Microsoft also released MAI Image 2.7, a new image-generation and editing model, along with a cheaper MAI Image 2.7 Flash version. Microsoft says MAI Image 2.7 ranks first on Artificial Analysis's image editing leaderboard, ahead of GPT Image 2.6 Sunburst, at a fifth of the price. The pricing is listed at about $35 per 1,000 images for the full model and about $10 per 1,000 for the Flash version. The lead, however, is narrow: 4 Elo points on the leaderboard as of 16 November 2026, a margin small enough that it could change with the next round of testing. Both models are in public preview in Microsoft Foundry, and the full model can also be tried in MAI Labs, a limited preview for personal, non-commercial use.

MAI Transcribe 2 and MAI Realtime 1 cover voice. Microsoft describes Transcribe 2 as its best transcription model, and Realtime 1 as its model for real-time voice interaction. Fewer details were published about either model at launch than about the coding and cyber models, and no benchmark figures for them were prominent in the announcement coverage. They appear to be supporting players in the family: useful for building voice features into Microsoft products, but not the reason most people will notice this launch.

MAI Decision 1.1, which Microsoft confirmed is generally available, is a different kind of model again. Instead of writing text or code, it scores answers — judging which of several possible responses is best. That is the kind of model that sits behind the scenes in an AI system, routing work and checking quality, and its availability signals that Microsoft is building the full infrastructure of an AI platform, not just a set of chatbots.

Taken together, the family covers the main jobs a business might want AI to do: write code, find security flaws, make images, transcribe speech, talk in real time and judge answers. The unifying idea is the one Suleyman put to Axios: specialising lets Microsoft "cram more tokens and knowledge and expertise into the same amount of parameters." A smaller model that knows one domain deeply can, in his argument, match a much larger general-purpose model on that domain's tasks while costing far less to run.

Why Microsoft Is Doing This Now

To understand the launch, it helps to understand Microsoft's position. The company is OpenAI's largest commercial partner and a major investor, and its Copilot products have run largely on OpenAI's models. That relationship has made Microsoft enormously successful in AI, but it has also left Microsoft dependent on a supplier that is also a competitor, and exposed to OpenAI's pricing. Building its own models is a way to reduce that dependence and to capture more of the value itself.

It is also a response to a change in the market. For the past two years, the AI industry has been dominated by a race to build ever-larger general-purpose models. That race is expensive: training and running frontier models costs billions of dollars, and only a handful of companies can afford it. Microsoft's argument is that the race is changing. As models become good enough for everyday work, the competitive question shifts from "which model is smartest?" to "which model is cheapest per task?" — and that is a question a company with Microsoft's cloud infrastructure is well placed to answer.

The strategy also fits with what Microsoft's own customers are telling it. Businesses that use AI coding assistants at scale care deeply about the cost per task, because they run millions of tasks. A model that is 5% of the price of the market leader, even if it is 4% worse on a benchmark, is an easy decision for routine work. The same logic applies to image generation and transcription, where volume is high and the quality bar is lower than for frontier research.

There is a wider industry context, too. Google has been pushing its own cheaper models — our recent coverage of Gemini 4 Hydrogen pricing and benchmarks shows the same pattern of smaller, cheaper models aimed at everyday work — and Alibaba and others have launched competing models in recent weeks. The era in which a single flagship model dominated every task is giving way to a market of many specialised models competing on price. Microsoft's launch is the biggest entry yet into that market.

None of this means Microsoft is abandoning OpenAI. The two companies remain deeply intertwined, and Microsoft continues to offer OpenAI's models alongside its own. What has changed is that Microsoft now has a credible alternative to offer its customers — and a negotiating position it did not have before.

The Outage Question: What AI Did and Did Not Do

Any Microsoft AI story published this week has to address the elephant in the room: the global Microsoft 365 outage of 12–14 November 2026, which knocked out Outlook, Teams and Copilot for millions of users for almost 24 hours, just days before the Ignite launch. As the outage unfolded, theories spread online that it had been caused by a cyberattack, or by an AI agent going rogue, or by AI-generated code pushed into production.

Microsoft has rejected those theories directly. In a statement posted on 13 November 2026, the company said: "This was a technical issue not related to cyber or AI / agents. The incident was introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change." Charles Lamanna, Microsoft's executive vice president for Copilot, Agents and Platform, confirmed there was "no indication this is the result of malicious or cyber activity." The company has promised a full post-incident report once its review is complete.

The timing is awkward for Microsoft — an AI company launching new AI models days after a major outage that the public blamed on AI — but the evidence available points to a networking fault, not to the models themselves. Readers who want the full account can read our earlier article on the Microsoft 365 outage of November 2026. The outage is a reminder that the infrastructure AI runs on can fail in ordinary, unglamorous ways, and that not every technology failure is an AI failure.

What to Watch Next

The launch raises several questions that will only be answered in the coming months. The first is whether independent testing confirms Microsoft's benchmark claims. Microsoft chose the benchmarks, ran the tests and published the results. Outside evaluations — from independent labs, from developers using the models in real work, and from leaderboard operators — will show whether the price-performance advantage holds up outside Microsoft's own testing setup. The promotional pricing also matters: the real test of MAI Code 2's value will come after May 2027, when the price doubles and the comparison with GPT-6.1 Sol becomes much closer.

The second question is how OpenAI and Anthropic respond. Both companies have been moving towards cheaper models of their own, and both have enormous advantages in the hardest tasks, where Microsoft admits its models are not yet tuned. If they cut prices, Microsoft's advantage narrows. If they do not, Microsoft will keep winning the routine work — which is where most of the volume is.

The third question is about the cyber model. MAI Cyber 2 Flash is the most capable security model Microsoft has ever built, and it is being kept behind a velvet rope. How Microsoft decides who counts as a "verified defender", and whether the model stays inside Microsoft's products or eventually becomes more widely available, will say a lot about how the company balances commercial opportunity against the risk of putting powerful vulnerability-finding tools into wide circulation.

The fourth question is internal. Microsoft's own staff have pushed back against the pace of the company's AI drive in recent months, and the departure of senior AI researchers — including the co-founder of Microsoft's AI Frontiers lab — has raised questions about whether the company can keep the talent it needs to sustain this strategy. Building a model family is one thing; keeping the team that built it is another. Our earlier reporting on Microsoft staff pushback against the AI drive and on John Langford's departure from Microsoft covers that tension in more detail.

For developers, the practical next step is simple: MAI Code 2 and Code 2 Flash are already in GitHub Copilot, so anyone with a Copilot subscription can try them in the model picker. The honest way to judge them is to run them on your own work — your own codebase, your own tasks — and compare the results and the bill against the models you use today. Benchmarks are a starting point, not a verdict. The price is real, the performance is Microsoft's claim, and the gap between the two is where the truth will be found.

What We Could Not Verify

  • Independent benchmark results for MAI Code 2. All performance and price comparisons published at launch come from Microsoft. No independent lab had published its own test of the models at the time of writing, so the headline figures (98% of GPT-6.1 Sol at 40% of the cost; 96% of Claude Opus 5.5 at 5% of the cost) are Microsoft's claims, not verified facts.
  • The full price after the promotional period. Microsoft's coding prices are promotional and due to double after 17 May 2027. The exact post-promotion pricing structure beyond that date has not been published in full, and the per-task costs at full price are estimates derived from Microsoft's own figures.
  • Details of MAI Transcribe 2 and MAI Realtime 1. Microsoft published far less detail about its voice models than about the coding and cyber models, and no benchmark figures for them were prominent in the launch coverage. Their capabilities and pricing are therefore described here only in general terms.
  • The full post-incident report on the November 2026 outage. Microsoft has said the outage was a network fault unrelated to AI, and has promised a full report once its review is complete. That report had not been published at the time of writing.

Sources

Similar Posts