MAI-Image-2

What it is

MAI-Image-2 is a text-to-image generation model from Microsoft AI. It is the second generation of the company's image model, and it was made available to developers on 2 April 2026 in Microsoft Foundry and the MAI Playground^[1]. A developer writes a text prompt and receives a generated image in return. Pricing at launch started at $5 per million text input tokens and $33 per million image output tokens^[1][2], so the cost scales with both the written prompt and the picture that comes back.

Where it came from

The model is the successor of MAI-Image-1^[3]. Microsoft said it debuted as a top-3 model family on the Arena.ai text-to-image leaderboard^[1]. The developer release came after the model was already running in production: Microsoft said the three models it shipped on 2 April 2026 were already powering Copilot, Bing and PowerPoint before they were offered to developers^[1]. For MAI-Image-2 specifically, the rollout to Bing and PowerPoint happened in phases, and the model was deployed in Copilot^[1].

What it is good at

Microsoft cited natural lighting, accurate skin tones and texture, and legible in-image text as the model's strengths^[1]. The first two are familiar goals for image generators, which are often judged on whether a face or a surface looks like something a camera could have recorded. Legible text is a harder problem: a model has to place letters that stay readable rather than dissolving into decorative shapes, which matters for anything resembling signage, packaging or a poster.

Speed was the other selling point. Based on production traffic, Microsoft said the model generates at least twice as fast as before on Foundry and Copilot at similar quality^[1]. Faster generation changes how the tool gets used in practice, since a person iterating on a prompt can try more variations in the same amount of time.

Release process and price

The company published model cards for the model and had it red-teamed before release^[1]. Red-teaming means deliberately probing a system for harmful or unintended outputs before it reaches users, and the published card documents what the model is and how it behaved in testing. The launch price, from $5 per million text input tokens and $33 per million image output tokens^[1][2], puts the model in a per-use billing model rather than a flat subscription, which lets developers estimate costs from the volume of prompts and images they expect.

Partners

Advertising group WPP was named as an early enterprise partner^[1]. WPP global chief creative officer Rob Reilly said the model responds to the nuance of creative direction and respects the craft of producing campaign-ready images^[1]. That is a statement about how usable the output is for commercial work, where an image often has to survive client review rather than simply look striking on its own.

Where it sits

MAI-Image-2 arrived as part of a group of three models Microsoft shipped on the same day, all of them already embedded in the company's consumer and productivity products before developers could call them^[1]. It was followed on 2 June 2026 by MAI-Image-2.5 and MAI-Image-2.5-Flash, which Microsoft said surpass the Arena score of Nano Banana Pro^[3]. Read together, the releases describe a fast cadence: a second-generation model in April, a refinement and a lighter-weight variant two months later, each measured against the same public leaderboard.