Updates·October 28, 2026, 04:25

Microsoft's new image editing AI lands on the podium in two leaderboards

AI-generated from the sources listed, with automated source checks.

Microsoft AI has released MAI-Image-2.7 in preview. The model edits images based on text instructions and ranks No. 2 or 3 on two leaderboards, depending on which one you look at.

Microsoft AI released MAI-Image-2.7-preview on October 27, 2026. The model is built to edit existing images based on text instructions. In other words, you write what needs to change, and the model adjusts the image.

According to all-ai.de, it can change the style or lighting of a scene, fix text and symbols in images, and restore older photos. It takes into account how objects, lighting and surroundings fit together.

Two leaderboards, two rankings

On Artificial Analysis' AA-Image-Editing v2.1 leaderboard, the model is tied for second place with 1,193 points. GPT Image 2.5 Sunburst leads with 1,199 points, while GPT Image 2.5 Flare trails with 1,182.

On Arena's Image Edit leaderboard, the model debuted at No. 3 with 1,470 points (±10). GPT-Image-2.5 Sunburst is first with 1,515, and GPT-Image-2.5 Flare is No. 2 with 1,483. GPT-Image-2 (Medium) trails with 1,460. Its predecessor MAI-Image-2.6 scored 1,425 (±5), which puts the new model 45 points ahead.

Microsoft AI itself writes on its blog that the model debuts above GPT Image 2.5 Flare at Artificial Analysis and at No. 3 on Arena.

Better than its predecessor, but not at everything

According to all-ai.de, the model beats MAI-Image-2.6 in all ten use cases examined and seven types of image editing. The improvements are largest for scenes, style, text, symbols and restoration.

The analysis site ExplainX cites launch reports and writes that the model leads in four of seven editing categories: scene and style, text and symbols, reasoning-based edits, and enhancement and restoration. It reportedly lags in composition, object editing and identity preservation. Scalevise, on the other hand, reports a big improvement in composition and framing, so the sources do not entirely agree.

ExplainX also warns that a ranking is not a verdict. Votes on what people prefer do not ensure that the result is correct. Preview models can change before the final release.

Where can you try it?

The model is in preview and can be tried in MAI Playground. According to all-ai.de, it is only for image editing, and text-to-image is planned for later. Scalevise writes that in preview it can also be tried via Arena and Artificial Analysis.

Pricing and API access have not been disclosed. The same goes for a broader rollout of 2.7. MAI-Image-2.6 and 2.6 Flash are now generally available in Microsoft Foundry.