Microsoft's MAI-Image-2.7 Climbs to No. 2 on Image-Editing Benchmark
Microsoft AI's newest image editing model climbs to #2 on the Artificial Analysis leaderboard, trailing only OpenAI's GPT Image 2.5 Sunburst and leading four categories.
- Microsoft's MAI-Image-2.7-Preview ranks #2 on AA-Image-Editing v2.1, behind only GPT Image 2.5 Sunburst (max).
- Beats predecessor MAI-Image-2.6 on all 7 editing actions, leads on 4 of them.
- Biggest Elo jumps: Composition & Framing (+57), Text or Symbol Edits (+49), Scene & Style Edit (+48).
- Ranks #1 on Live-Action Film, Consumer, Frontier, and UI/UX Design workloads.
- Available today in MAI Playground for image editing only; text-to-image coming soon.
- GPT Image 2.5 Sunburst still leads on Object-Level and Identity Preserving Edits.
Microsoft’s MAI-Image-2.7-Preview ranks second on Artificial Analysis’s AA-Image-Editing v2.1 leaderboard, behind OpenAI’s GPT Image 2.5 Sunburst (max) and ahead of MAI-Image-2.6. The task-level results identify several workloads where Microsoft’s latest in-house editor leads the benchmark, alongside areas where Sunburst retains an advantage.
Microsoft previously released three MAI models through Microsoft Foundry: MAI-Image-2, MAI-Voice-1, and MAI-Transcribe-1. The 2.7 preview extends that first-party model portfolio, although Microsoft has yet to announce a public API for the new image editor.
Four category leads, three gaps
Artificial Analysis divides image editing into seven actions, which describe the requested operation, and ten use cases, which represent broader workloads. MAI-Image-2.7-Preview improves on version 2.6 in all seven action categories and ranks first in four.
| Leading action | Typical tasks |
|---|---|
| Scene & Style Edit | Relighting, restyling, background changes, and broader visual treatments |
| Text or Symbol Edits | Typography, logos, labels, and other text embedded in an image |
| Reasoning-Based Edit | Changes that require inferring intent or relationships within a scene |
| Enhancement & Restoration | Repairing, cleaning, or improving existing images |
The largest Elo gains over MAI-Image-2.6 appear in Composition & Framing at 57 points, Text or Symbol Edits at 49 points, and Scene & Style Edit at 48 points. Elo measures relative benchmark position, so those changes represent rating-point gains rather than percentage improvements.
Sunburst still leads Composition & Framing, Object-Level Edit, and Identity Preserving Edit. Those categories cover important production tasks such as repositioning subjects, replacing individual products, and modifying portraits without changing a person’s recognizable features.
These rankings describe performance on the benchmark’s prompts and scoring process. They do not establish production latency, failure rates, moderation behavior, API stability, or cost, all of which can change the preferred model for an application.
Film, editorial, and complex edits stand out
Across the ten use-case categories, MAI-Image-2.7-Preview ranks first in four. Sunburst leads the remaining six, with only a narrow gap between the models in UI/UX Design.
| Leading use case | Examples |
|---|---|
| Live-Action Film | Cinematic scenes, film stills, storyboards, and character sheets |
| Consumer | Book covers, stock images, and editorial illustrations |
| Frontier | Single requests that combine several editing capabilities |
| UI/UX Design | Interface mockups, screens, and visual product concepts |
The category split makes MAI-Image-2.7-Preview a strong candidate for testing on stylized film work, editorial assets, interface design, and multi-step edits. Sunburst remains the benchmark leader for workflows that depend heavily on object replacement, precise framing, or identity retention.
Playground access, with no 2.7 API yet
MAI-Image-2.7-Preview is available in MAI Playground, Microsoft’s browser-based testing environment. Access is free with a Microsoft account. The preview currently handles image editing, and Microsoft says text-to-image support will follow.
| Access | MAI Playground |
|---|---|
| Current mode | Image editing |
| Public API | No 2.7 endpoint announced |
| Pricing | No public 2.7 API pricing |
| Weights | Closed, with no self-hosting option |
| Regional access | The Playground has historically been limited to US users; availability may vary |
Earlier releases reached Microsoft Foundry through a multipart form-data image-editing API that accepted as many as five reference images, according to Microsoft’s earlier API guide. Those models charged separately for text input, reference-image input, and image output. Microsoft may change the interface, limits, or billing model before releasing a 2.7 endpoint.
Five releases in roughly a year
- October 2025: MAI-Image-1
- March 2026: MAI-Image-2
- June 2026: MAI-Image-2.5
- August 2026: MAI-Image-2.6
- Latest release: MAI-Image-2.7-Preview
Microsoft’s work on a first-party image stack strategy gives the company direct control over model updates, capacity, safety policies, product integration, and unit costs. It can also supply models for products such as Bing Image Creator, PowerPoint, Copilot, and OneDrive while reducing dependence on licensed image systems.
The rapid release sequence has produced measurable benchmark gains with each version. MAI-Image-2.7-Preview is the first model in the family to sit directly behind OpenAI’s leading editor on this independent leaderboard.
Build a task-specific evaluation
Teams evaluating image editors should compare both models on representative assets because aggregate Elo ratings can hide application-specific failures. A useful test set should cover the operations that appear most often in production and retain identical prompts, source images, and output settings across models.
- Measure instruction adherence: Check whether each requested change appears without unwanted alterations elsewhere.
- Test identity and object consistency: Include faces, branded products, recurring characters, and multiple reference images.
- Inspect embedded text: Score spelling, typography, placement, and consistency across repeated runs.
- Record operational behavior: Track generation time, rejected requests, retries, output variability, and failure rates.
- Recalculate production cost: Compare pricing after Microsoft publishes the 2.7 API and billing details.
Without a public 2.7 endpoint, production adoption remains pending. Playground tests can establish whether its strengths justify a later integration, while task-specific routing may suit teams whose workloads span both Microsoft’s leading categories and Sunburst’s stronger object and identity edits.
Comments
