Microsoft's MAI models chase cheap tokens, not bragging rights

Microsoft now offers more than 11,000 models through its Foundry platform, and until recently not one of them was a home-grown frontier model. That gap has closed. The company's Microsoft AI (MAI) program has begun rolling out its own foundation models, built around a single argument: satisfactory results at a lower token cost than Anthropic or OpenAI.

The backstory matters here. When Microsoft first deployed AI models in what became Foundry, it leaned on its OpenAI partnership — OpenAI trained on Azure, and Microsoft became the first commercial partner to run those models. Azure OpenAI has been a successful product for both businesses. But as more organizations gained access, Microsoft's catalogue expanded to include models from Anthropic, Hugging Face, Meta, Mistral and DeepSeek, alongside its own Cognitive Services models for computer vision and text processing, and the Phi series of lightweight edge models.

What does Microsoft actually bring to the table?

Microsoft Research has spent much of the last decade in deep learning, and its models regularly top benchmark tables. What it never did was chase the general-purpose approach of Anthropic, OpenAI or Google. It built focused models instead. MAI continues that pattern: models aimed at cybersecurity, coding and transcription, with a tuning service that lets customers build custom versions on their own data.

The timing is deliberate. The days of free and low-cost tokens are gone, as hyperscalers look to recoup multi-billion-dollar investments in AI data centers. So-called "tokenmaxxing" — dumping massive data sets into LLMs and hoping for results — is no longer economically viable. MAI describes its approach as "humanist", with a stated focus on safety and environmental impact, using models to choose models and dedicated hardware that processes tokens for fewer watts.

How do you deploy MAI-Thinking-1?

Most MAI models are available through Foundry. Two are reserved: MAI-Cyber-1.1-Flash powers Microsoft's MDASH security system, and MAI-Code-1.1-Flash runs inside GitHub Copilot. For complex applications, there is MAI-Thinking-1, the program's reasoning model. Microsoft says it runs with a much smaller footprint than competing models while still delivering expected results, and the company publishes benchmarks that favor some bigger models from Anthropic and OpenAI.

The pitch is not absolute quality but economics. MAI-Thinking-1 is called through the existing Foundry Azure OpenAI APIs, so migration is modest: deploy the model to your Foundry instance, change the API call to the model's endpoint URL, and start sending messages. You can cap the number of tokens per request to keep application budgets in check.

The model supports function calls for agent building, and it can save encrypted reasoning chains between requests, letting an agent move quickly to the next response in a chain. Like other large models, it can stream chunks back to the client rather than delivering a complete chat object — useful for user experience, unnecessary in a long-running workflow. There, Microsoft's new decision-scoring model can serve as an agent router, deciding which steps run and which models each step uses.

Why the Flash models matter

The new Flash models run locally or in the cloud, offering similar fidelity to cloud-only models at significant savings. Local inference on NVIDIA RTX-class hardware draws considerably less power than data center gear. The catch is memory: MAI-Code-1.1-Flash needs more than 120GB of RAM, which rules out Copilot+ PCs on Arm silicon. The NVIDIA RTX Spark-powered Surface Ultra, configured with at least 128GB of RAM, should handle it — though buyers will need to weigh agent usage against the price of such a device.

MAI-Cyber-1.1-Flash has a different job. It offloads 90% of the tasks that previously went to much larger, more expensive models in the MDASH platform, passing the remainder to more powerful models when it cannot cope. The 1.1 release is also 18% cheaper to run than 1.0.

What does Maia 300 change for inference costs?

The models sit on Microsoft's own silicon. The latest Azure Maia inference accelerators are optimized for Microsoft's models and for OpenAI, and the underlying dataflow architecture lets inference software place data where the accelerator accesses it most efficiently. Maia is no longer just a chip: it is a platform spanning rack, server and networking, using non-uniform memory access and what Microsoft calls a Software Defined Locally Accessed Dataflow Architecture, with a chiplet design that separates memory and I/O.

The numbers are concrete. Maia 300 boosts compute operations by 36%, delivers 2.5x improvements in high-bandwidth memory performance, and roughly doubles overall performance per watt. Each rack holds 96 accelerators, supports multiple jobs, and can shard work across racks — multi-tenant inference without blocking operations.

The verdict for buyers watching growing AI bills: you don't need the absolute best model, only one good enough and cheap enough for most applications. That is precisely what Microsoft is shipping.