Hot AI News

Microsoft’s seven new MAI models mostly sell on price, not on topping rivals

Microsoft's MAI group released seven models and the Maia 200 chip, arguing that agents running around the clock need frontier quality at a fraction of rivals' prices.

Microsoft’s seven new MAI models mostly sell on price, not on topping rivals
Image: Microsoft

Microsoft’s MAI group released seven new models at Ignite, covering code, security, image editing, real-time voice, transcription and decisions. The pitch is cost: MAI Code 2 reaches 98% of GPT-6.1 Sol’s coding quality at about 40% of the price, according to Microsoft.

The company also said its second-generation Maia 200 accelerator is now in production, running its largest models, with 1.7x more tokens per watt.

Header artwork for the MAI Code 2 launch at Microsoft Ignite
Image: Microsoft

Seven models, each with a price claim

Microsoft framed the release around “token spend skyrockets” and a demand for models cheap enough to run agents continuously on real engineering work. Almost every claim is a ratio against a named rival:

  • MAI Cyber 2 Flash matches GPT-6 Astra’s quality on the CyberGym-E2E benchmark at under 2% of the cost per task.
  • MAI Code 2 reaches 98% of GPT-6.1 Sol’s coding quality at about 40% of the price, and 96% of Opus 5.5’s performance at 5% of the price.
  • MAI Code 2 Flash is described as comparable to GPT-6 Luna at 65% of the price, and to Haiku 5.5 at 20% of the price.
  • MAI Image 2.7 ranks first on Artificial Analysis image editing and, Microsoft says, beats GPT Image 2.6 Sunburst (max) at one-fifth of the price.
  • MAI Realtime 1 is said to outperform GPT-Realtime 2.1, with prices starting at $0.01 per audio minute.
  • Dragon Copilot Transcribe runs on MAI Transcribe 2, which Microsoft calls the fastest and most accurate multilingual transcription model. The company says the product already supports 10 million patient encounters monthly and has cut multilingual errors in half.
  • Decision 1.1, Microsoft’s newest decision model, is described as 50x faster and 79x cheaper than GPT-6.1 Sol.

The announcement gives few absolute prices. The one it states clearly is the $0.01 per audio minute starting rate for MAI Realtime 1.

Sample image from the MAI Image 2.7 model
Image: Microsoft
Advertisements

Who can use these models today

Access differs by model. MAI Code 2 and MAI Code 2 Flash are in GitHub Copilot. In our earlier report on the coding models, MAI Code 2 is open to Pro, Pro+, Max, Business and Enterprise plans, and Flash also reaches Copilot Free and Student. Dragon Copilot Transcribe is live inside Dragon Copilot. The page lists no regions, and availability for the other models should be checked in Microsoft’s own documentation.

Cost per task becomes the contest

The release shifts the argument from raw leaderboard position to price at a given quality level. That matters for agent workloads, where a model may be called thousands of times a day and a 20x price gap decides whether a job is affordable.

Artwork for the MAI Realtime 1 voice model
Image: Microsoft

Two details temper the headline figures. First, GitHub’s own coverage says Microsoft’s cost numbers for the coding models include a 50% promotional price, so ratios could narrow when promotions end. Second, the 96% and 98% quality figures come from Microsoft’s average over four coding benchmarks, run at maximum reasoning effort. In blind frontend tests with 258 annotators, reviewers preferred MAI Code 2 over GPT-6 Sol 64% of the time, per the same earlier report. All of this is Microsoft’s own testing; none of it has been independently verified.

Quality is also stated as a fraction of the leader, not a win. Microsoft is saying that a slightly weaker model at a small share of the cost is the better trade for sustained agent work, and it is making that case against OpenAI and Anthropic models it also distributes through its platforms.

The Maia 200 claim fits the same logic. More tokens per watt lowers serving cost on Microsoft’s side, which is what makes aggressive pricing sustainable. The source gives no baseline for the 1.7x figure, no price per token and no capacity numbers.

Advertisements

What the announcement leaves unproven

Several claims lack the detail needed to check them. The page gives no context windows, parameter counts or absolute token prices for the code and cyber models. For Decision 1.1, the 50x and 79x figures are measured against GPT-6.1 Sol, but the source does not say on which tasks. The transcription result of halved error rates does not state the earlier model or language set used as the baseline.

Customers who want to test the claims will need to run their own workloads, particularly for the Flash models, where “comparable performance” is a loose description.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like