news

Microsoft's MAI Image 2.7 lands second on Artificial Analysis image editing, six Elo behind first

The preview scores 1,193 Elo against GPT Image 2.5 Sunburst's 1,199, and the only three places to reach it are a playground and two leaderboards.

By Priya Shenoy · 28 October 2026

Microsoft AI released MAI Image 2.7 in preview on October 27 with a leaderboard placement and very little that a production team can act on. The model edits existing images but cannot generate one from a prompt, and the three places Microsoft names for trying it are MAI Playground, Arena and Artificial Analysis, which is one demo surface and two evaluation surfaces.

Six Elo separates the preview from first place on AA-Image-Editing v2.1

On Artificial Analysis's image editing board, MAI-Image-2.7-Preview entered at 1,193 Elo, second behind OpenAI's GPT Image 2.5 Sunburst at 1,199 and ahead of GPT Image 2.5 Flare at 1,182. Microsoft's previous model, MAI-Image-2.6, sits at 1,151 on the same board. That makes the generational gain 42 points and the gap to the top 6. Artificial Analysis called the result a clear step up from 2.6, which is a separate claim from being close to first, and a single chart carrying both numbers makes the two easy to blur together. A six-point spread on a preference-vote board is the kind of margin that has collapsed into a tie once confidence intervals are drawn on other image arenas.

First on four of seven editing actions, and four of ten use cases

The Artificial Analysis editing taxonomy splits into 7 editing actions and 10 use cases. By the evaluator's own account, 2.7 scores higher than 2.6 on all 7 actions and ranks first on 4 of them: scene and style edits, text or symbol edits, reasoning-based edits, and enhancement and restoration. The largest movements against 2.6 are composition and framing at plus 57 Elo, text or symbol edits at plus 49, and scene and style at plus 48. On the use-case side it ranks first for live-action film, consumer and frontier edits, and narrowly leads UI and UX design.

Microsoft's version of the same data is broader. Its post says 2.7 improves on 2.6 across all 10 use cases and all 7 editing actions evaluated by Artificial Analysis, that it reaches the category frontier in enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing, and that its strongest lead is 24 Elo in scene and style. Both statements can be true at once, but the per-category rankings are the ones that identify where the model actually leads.

Third on Arena, described as second among labs

Microsoft reports a No. 3 debut on Arena Edit and frames that as making Microsoft AI the No. 2 lab on that board, a reframing that converts a third place into a second by grouping models under their makers. The same arithmetic showed up when Canva Craft entered DesignArena fourth and was described as second among labs.

No generation mode, no endpoint, no published price

The announcement commits to text-to-image generation only as coming soon. It names no API, no output resolution, no reference-image limit and no price, which leaves the Elo figure as the single quantitative fact attached to the release. For anyone scoring image editors rather than buying them, the more interesting detail is that the model's entire public footprint on launch day consisted of a playground and the two boards that rank it, a pattern that makes leaderboard position the product rather than a measurement of it.

Sources: Microsoft AI, Artificial Analysis, Neowin.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.