Overview
MAI-Image-2.7 is Microsoft's image editing model, released in preview on October 27, 2026. It does one thing: take a source image and a text instruction, and return the edited image. Text-to-image is described as coming soon without a date.
On Artificial Analysis AA-Image-Editing v2.1 it scores 1193 Elo for second place, behind OpenAI's GPT Image 2.5 Sunburst at max. Microsoft says 2.7 improves on 2.6 across all ten use cases and all seven edit actions, reaching the frontier in four categories: restoration and enhancement, reasoning edits, scene and style, and text and symbols. The largest margin is in scene and style, 24 Elo ahead of the next model.
The model targets edits that are hard to get right: wholesale style changes, relighting, rewriting text and symbols inside a frame, repairing and enhancing footage, and compound edits that depend on understanding how objects, light and environment relate.
Key Features
- Editing only, no generation: Input is an image plus an instruction; output is the edited image. Restricting scope to existing photos, design files, screenshots and footage keeps the focus on fidelity.
- Frontier in four edit categories: Restoration, reasoning edits, scene and style, and text and symbols all reach the frontier. All ten use cases and seven actions improve on 2.6, so this is a broad lift rather than one standout trick.
- 24 Elo ahead on scene and style: That is the widest gap of the four. Style transfer is the edit most likely to damage object identity, which makes it the most useful of the gains.
- Editable text and symbols: Lettering on posters, interface screenshots and packaging has historically been the least controllable part of an edit. Listing it as a frontier capability means changing words semantically instead of repainting the region.
- Relighting that respects the scene: Microsoft highlights object, light and environment relationships, which in practice means keeping existing reflections and shadows consistent rather than compositing a new light layer.
- Three places to try it: Arena, Artificial Analysis and the MAI Playground all serve the model. Playground availability may be region-limited and Microsoft has not published the full list.
Use Cases
- Refreshing commerce and advertising assets by changing style and lighting while keeping the product shape and material.
- Replacing lettering on posters and packaging semantically rather than repainting the area.
- Restoring old or low-quality images with detail recovery that keeps the original look.
- Pitching film and video treatments by regrading live-action footage across eras and moods.
- Localising interface screenshots while preserving layout and typographic feel.
- Reasoning-style edits to a design, adjusting composition from a description instead of smudging pixels.
Pros
- 1193 Elo on AA-Image-Editing v2.1 for second place.
- All ten use cases and seven edit actions improve on the previous generation.
- Scene and style leads the next model by 24 Elo.
- Handles text and symbols inside the frame, historically the hardest part.
- Relighting accounts for object, light and environment relationships.
- Available to try in Arena, Artificial Analysis and the MAI Playground.
Pricing
In preview. Microsoft has not published a public API, pricing or weight release. Access is limited to Arena, Artificial Analysis and the MAI Playground, and playground availability may vary by region. Commercial terms will follow the general availability release.
Summary
The generational bet is narrow: skip text-to-image and get editing right first. For anyone reworking existing assets the payoff lands in two specific abilities, changing style without damaging the subject and rewriting in-frame text semantically. The constraints are equally clear: no API and no weights during preview, with region limits on the playground, so production integration waits for the full release.
Version History
- MAI-Image-2.7-Preview released: second on the image editing leaderboard (2026-10-27): Microsoft released the MAI-Image-2.7-Preview image editing model, scoring 1193 Elo on Artificial Analysis AA-Image-Editing v2.1 for second place. Against 2.6 it improves in all ten use cases and seven edit actions, reaching the frontier in restoration, reasoning edits, scene and style, and text and symbols, with scene and style leading the next model by 24 Elo. The preview handles image plus instruction editing only, with text-to-image announced as coming soon. It is available in Arena, Artificial Analysis and the MAI Playground, with no API, pricing or weights published yet.