Timeline

Microsoft AI releases seven models, including MAI Code 2

Microsoft reports MAI Code 2 reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price; Axios noted the comparisons use selected benchmarks and some prices are promotional.

  • Models & capabilities
  • Labs & people
  • Notable

Microsoft AI, the company’s in-house model group, released what it called “a family of seven new world class models” at its Ignite conference, led by MAI Code 2, a coding model now in GitHub Copilot. The announcement stresses price-performance: models that Microsoft says approach the best outside systems on specific tasks at a fraction of the cost.

The lineup, as Microsoft describes it, is MAI Code 2 and a smaller MAI Code 2 Flash; MAI Cyber 2 Flash for security work; MAI Image 2.7; MAI Realtime 1, a speech-to-speech model starting at $0.01 per audio minute; Dragon Copilot Transcribe, built on a model called MAI Transcribe 2; and Decision 1.1, a routing model. Axios’s headline counted six models rather than seven.

Every comparison is Microsoft’s own. It says MAI Code 2 gives 98% of the performance of GPT-6.1 Sol at 40% of the price, and 96% of Claude Opus 5.5’s at 5% of the price, averaged over four coding benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1). It says MAI Cyber 2 Flash beats GPT-6 Astra on the CyberGym-E2E benchmark at under 2% of the cost per task, and that MAI Image 2.7 ranks first on Artificial Analysis’s image-editing board. Axios added a caveat from its own reading: the comparisons “rely on selected benchmarks, and some pricing advantages reflect promotional discounts”. Microsoft AI chief executive Mustafa Suleyman told Axios the approach packs more knowledge into a given number of parameters, at the cost of weaker performance outside the target tasks such as advanced mathematics.

Microsoft also announced MAI Licensed Weights, under which enterprise customers can customise a frontier MAI model on their own data and run it in their own Azure environment. Axios reported that Suleyman declined to give a date for a general-purpose frontier model, saying his team was making progress on pre-training and reinforcement learning.

The release shows Microsoft building a model line of its own alongside its use of outside models; OpenAI released GPT-6.2 Astra and Sol the same day. The first test is how much Copilot traffic moves to MAI Code 2, a figure Suleyman declined to forecast.