Industry Impact2 min read

Microsoft's Cheap AI Models Make Checking Everything Affordable

By , Senior AI ConsultantPublished

Microsoft's new AI models do everyday work for a small fraction of rival prices, and like cheap light two centuries ago, that should make companies check and sort far more than they used to.

At its Ignite conference in November, Microsoft released a set of its own AI models for coding, security, images, speech and sorting requests. It priced them well below the OpenAI and Anthropic models that do similar work inside its products. On Microsoft's own tests, its coding model reaches 96% of Claude Opus 5.5's quality at 5% of the cost, and its security model costs less than 2% as much per task as GPT-6 Astra.

Cheap light changed what people did with light

Economists have measured the price of light by the hours of work a person needed to buy it. 1,000 lumen-hours, about what a 60-watt bulb gives in a little over an hour, took roughly five hours of work in 1800 and less than half a second of work in 1992.

The savings did not stay in household budgets. They went into uses that nobody had counted as lighting before: street lamps, shop windows at night, factories that ran after dark, reading in bed.

Security shows the same arithmetic

Security is where the price gap is widest. On a standard test of finding software vulnerabilities, Microsoft's new model finishes a task for about $0.45, while GPT-6 Astra costs $34.38 and solves slightly fewer of them. A security team with $1,000 a month gets about 29 tasks from the expensive model and about 2,200 from the cheap one, if real work costs anything like the test.

At 29 tasks a month, a team picks the few systems that deserve a deep look and leaves the rest alone. At 2,200, it checks every system each time something changes. That is a different kind of security, and it comes from the price, not from a smarter model.

You cannot buy this model on its own. It runs only inside Microsoft's security products, and customers of its MDASH scanning platform get it from mid-December.

The human decision moves, it does not disappear

The same shift reaches ordinary offices. Microsoft's decision model sorts and routes requests, picks only from actions a developer has approved, and reports how sure it is. It costs $0.042 per million input tokens and nothing for output, and it is available now.

An operations manager who builds a small tool on it could have every supplier email and customer complaint sorted on arrival. Anything the model is unsure about goes to a person. When sorting costs almost nothing, the scarce thing is a person's attention, and the confidence score tells the manager where to spend it.

What stays expensive

These models are built for everyday work. Microsoft says its coding models are not yet tuned for hard research or long, many-step jobs, so the larger models still handle those. That is the same logic as matching the model to the task, taken down to a much lower price. The low price of MAI Code 2 in GitHub Copilot is also a promotion that runs until May 17, 2027, after which it doubles.

Microsoft's own pitch is that customers' token spending is climbing fast because they want agents running around the clock, and cheap light would have predicted that. Cheaper use bought much more use. A company that pays 50 times less per task and runs 100 times as many tasks pays twice as much as before, and gets far more checked, sorted and caught for it.

Share this

STAY INFORMED

Get AI intelligence like this delivered to your inbox.

Free forever · Unsubscribe anytime


You May Also Find Valuable