
Microsoft Prices Its Own AI Models Under Its Partners'

Microsoft used its Ignite conference to put seven models of its own on sale, and the pitch is arithmetic rather than capability. MAI Code 2, the flagship coding model, is claimed to reach 98 per cent of GPT-6.1 Sol's quality at about 40 per cent of the price, and 96 per cent of Claude Opus 5.5's at five per cent.

The reason is where the models sit. Microsoft AI chief Mustafa Suleyman frames the family as a system that tunes quality and price together rather than chasing the frontier on every task, and the company says the coding models are deliberately not tuned for fundamental science or very long agentic work, because customers would pay for that capability on every routine fix.
And the same logic carries into security. MAI Cyber 2 Flash has 12 billion active parameters and Microsoft claims it edges past GPT-6 Astra on the CyberGym test, which runs a security task from start to finish, at under two per cent of the cost per task, with output tokens priced 86 per cent below the model it replaces in MDASH.
The catch is availability. Cyber 2 Flash is not sold on its own. It ships inside Microsoft's security products, starting with MDASH for eligible customers from next month, so a defender who wants it has to be inside that stack.
The rest of the lineup is broader than coding. MAI Image 2.7 takes the top spot on Artificial Analysis for image editing at a fifth of the price of GPT Image 2.6 Sunburst, MAI Realtime 1 is Microsoft's first speech to speech model from a cent per audio minute, and MAI Transcribe 2 now handles ten million patient encounters a month through Dragon Copilot.
And the strategic half of the announcement is about control rather than price. MAI Licensed Weights lets a customer take a frontier model, tune it on their own data, and run it in their own Azure environment, which is aimed squarely at companies that will not put their working knowledge into someone else's model.
The claims need context. Every figure is Microsoft's own testing measured against rivals' published benchmark results rather than one independent run, and the benchmarks are its choice, so the percentages describe a favourable comparison rather than a settled ranking. The company's Maia 200 accelerator is now in production on its largest models at 1.7 times more tokens per watt, which is where part of the cost story comes from.
Seven MAI models launched at Ignite in San Francisco
MAI Code 2 claims 98 per cent of GPT-6.1 Sol quality at 40 per cent of the price
MAI Cyber 2 Flash is limited to Microsoft security products, starting with MDASH
MAI Image 2.7 ranks first on Artificial Analysis for image editing
MAI Licensed Weights lets customers run tuned models in their own Azure environment
Source: Microsoft AI


