Microsoft's MAI-Image-2.7-Preview Takes #2 on AA-Image-Editing v2.1

Where MAI-Image-2.7-Preview lands on the editing leaderboard

Microsoft AI shipped MAI-Image-2.7-Preview two months after August's MAI-Image-2.6, and the independent evaluation from Artificial Analysis puts concrete numbers on the jump. On the AA-Image-Editing v2.1 benchmark, the new preview model takes the #2 spot across the board, sitting just behind OpenAI's GPT Image 2.5 Sunburst (max). In the wider Artificial Analysis Image Arena, it holds the #3 position with an Elo rating of 1470.

What makes the release stand out is that the gains are not confined to a single trick. Artificial Analysis grades image editors across seven distinct editing actions, and MAI-Image-2.7-Preview scores higher than MAI-Image-2.6 on all seven. Rather than trading off fidelity in one area to win another, Microsoft's update moves the entire baseline up while claiming the #1 rank on four of the seven action categories.

Seven editing actions and the three biggest Elo jumps

Looking inside the category breakdown shows where Microsoft focused its training. The four actions where MAI-Image-2.7-Preview ranks #1 outright across all tested models are Scene & Style Edit, Text or Symbol Edits, Reasoning-Based Edit, and Enhancement & Restoration. Those four cover the tasks that usually break older diffusion editors: rewriting signs without garbling typography, inferring what needs to change in a scene from an indirect instruction, and cleaning up degraded photographs without inventing plastic textures.

When measured directly against MAI-Image-2.6, three Elo deltas explain most of the climb up the table. Artificial Analysis also groups models by practical production domain, ranking MAI-Image-2.7-Preview #1 for Live-Action Film, Consumer, and Frontier edits.

• Composition & Framing (+57 Elo)

  The single largest leap over MAI-Image-2.6, covering reframing, subject repositioning, and aspect-ratio adjustments that keep perspective coherent.

• Text or Symbol Edits (+49 Elo)

  Replacing, inserting, or correcting lettering and logos inside an existing image while matching the surface lighting and font weight.

• Scene & Style Edit (+48 Elo)

  Relighting environments, swapping weather or time of day, and restyling a frame without losing the identity of the foreground subject.

Testing it in MAI Playground before text-to-image arrives

Right now, MAI-Image-2.7-Preview is scoped strictly to editing existing images. Microsoft AI has put the model live inside MAI Playground, where users can upload a source image in the browser and run editing instructions without installing local software. Text-to-image generation from scratch is not enabled on this preview checkpoint yet, though Microsoft states that text-to-image support is coming soon.

For teams comparing image editing APIs or deciding whether to switch from GPT Image 2.5 or MAI-Image-2.6, the full breakdown and side-by-side output samples are posted on artificialanalysis.ai inside the Image Arena. Running the same framing, signage, and live-action restoration prompts in MAI Playground gives a quick read on how those +48 to +57 Elo gains translate to your own assets.