Microsoft just told enterprises they no longer need to pay frontier prices for everyday AI work, and it brought its own models to prove the point.
At its Ignite conference in San Francisco on Tuesday, November 17, Microsoft unveiled a new family of in-house AI models covering coding, cybersecurity, image editing, real-time voice, transcription and model routing. The headline claim is aggressive: the company says its new MAI Code 2 model delivers close to the performance of the best coding systems from OpenAI and Anthropic at a small fraction of the price. The launch is Microsoft's clearest signal yet that it intends to compete directly with the AI labs it has long relied on, rather than simply reselling their technology.
What Microsoft announced
In its official Ignite announcement, Microsoft AI described the release as "a family of seven new world class models" built for price-performance. The lineup includes MAI Code 2 and the lighter MAI Code 2 Flash for software development, MAI Cyber 2 Flash for security work, MAI Image 2.7 for image generation and editing, MAI Realtime 1 for live speech-to-speech conversations, MAI Transcribe 2 (which powers Dragon Copilot Transcribe in healthcare), and Decision 1.1, a model that routes requests so that larger, more expensive models are only used when needed. The company also said Maia 200, its second-generation in-house AI accelerator, is now in production running its largest models.
Outlets counted the launch slightly differently: Axios reported six new models spanning coding, cybersecurity, image generation and voice, while Microsoft's own post groups them as seven. Either way, the breadth of the release is the point.
The numbers behind the pitch
Microsoft's benchmark claims are specific. According to the company, MAI Code 2, averaged across four coding benchmarks that reflect everyday enterprise tasks seen in GitHub Copilot (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), reaches 98% of the performance of OpenAI's GPT-6.1 Sol at about 40% of the price, and 96% of Anthropic's Claude Opus 5.5 performance at just 5% of the price. In blind side-by-side frontend tests, Microsoft says developers preferred MAI Code 2's output over GPT-6 Sol 64% of the time.
The smaller MAI Code 2 Flash is pitched for high-volume work such as code explanations, targeted edits and routine fixes. Microsoft says it matches GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20% of the price. Both coding models are available in GitHub Copilot now, and Microsoft's coding model post says they will soon come to third-party platforms including OpenRouter and Vercel.
Other claims include MAI Cyber 2 Flash matching GPT-6 Astra's quality at less than 2% of the cost per task on the CyberGymE2E benchmark, MAI Image 2.7 taking the top spot on Artificial Analysis's image-editing ranking at one-fifth the price of a leading OpenAI image model, and MAI Realtime 1 starting at $0.01 per audio minute. Microsoft says Decision 1.1 is used by 4,000 customers and is 50 times faster and 79 times cheaper than GPT-6.1 Sol, and that Maia 200 achieves 1.7 times more tokens per watt.
These are Microsoft's own figures. Axios noted that the comparisons rely on selected benchmarks and that some pricing advantages reflect promotional discounts, so independent testing will matter.
Why specialized models, and why now
The strategy is narrow by design. Microsoft openly says its coding models are "explicitly not yet tuned" for harder problems such as fundamental science research or very long-horizon agentic work. Instead, they target the bread-and-butter tasks that make up most enterprise usage.
Cost is the driver. Microsoft AI CEO Mustafa Suleyman said he keeps hearing from businesses that "token spend is going through the roof," according to Yahoo Finance. As companies move from chatbots to AI agents that run around the clock, every token adds up, and a model that is almost as good for a fraction of the price becomes very attractive.
Suleyman framed the approach as a repeatable system rather than a one-off. "This isn't just one model. It's not a flash in the pan," he said in an interview, according to Reuters, describing a "hill-climbing machine" that co-optimizes performance and price.
Security and control take centre stage
The launch lands amid heightened concern about AI agents behaving badly. Reuters reported that Suleyman said the recent cybersecurity reckoning, illustrated by OpenAI agents' hacking of the Hugging Face platform, had pushed him to focus on safeguards as well as performance, and that alignment was his top priority. Microsoft says it plans to offer its cyber model to a limited group of customers to help defenders cope with a surge in software vulnerabilities that AI tools are surfacing.
Microsoft also introduced what it calls MAI Licensed Weights, letting customers take a MAI model, customize it with proprietary data and run it inside their own Azure environment. The company cited AT&T, which it says built a custom MAI model for 5G network analytics that outperformed GPT-6.1 Sol at a far lower cost.
What it means for OpenAI, Anthropic and Copilot users
The most immediate effect may be inside Microsoft's own products. Because MAI Code 2 is available in GitHub Copilot, some coding workloads could shift away from outside models. Asked how much Copilot traffic Microsoft expects to move to its own models, Suleyman declined to give a target but predicted significant adoption, Axios reported. He also told Axios that Microsoft is making strong progress on a general-purpose frontier model, without offering a release timeline.
For developers and businesses, the practical takeaway is more choice and downward pressure on prices. For OpenAI and Anthropic, it is a reminder that one of their biggest distribution partners is now also a determined competitor. The next test is independent benchmarking and real-world adoption: whether teams using Copilot notice any drop in quality, and whether rivals respond with price cuts of their own.