Microsoft MAI-Image 2.7: an AI image editor ranked No. 2

what Microsoft announced on October 27

Microsoft AI released MAI-Image 2.7 in preview on October 27, 2026. For now it only edits existing images, and it debuts at No. 3 on the Arena Edit leaderboard and No. 2 on Artificial Analysis.

The announcement page is a two-minute read and the headline carries the news: MAI-Image 2.7 lands above GPT Image 2.5 Flare on Artificial Analysis and takes third place on Arena. Microsoft pitches it as its most precise editing model yet, built for the edits that are "hardest to get right". It lists four kinds of work: restyling or relighting a whole scene, changing text and symbols while keeping them part of the image, restoring or enhancing existing assets, and edits that depend on how objects, lighting and the environment affect each other.

First surprise: MAI-Image 2.7 does not generate images from a prompt. Microsoft says it is starting with editing, with text-to-image "coming soon" and no date attached. You can try it in three places: Arena, Artificial Analysis and MAI Playground. There is no API, no price and no general availability date.

We cover it here because editing is what most of our readers spend their time on: swapping the background of a product shot, rescuing bad lighting, fixing the text on a poster without regenerating the whole thing. A model that climbs to the top on exactly those jobs deserves a close look at its numbers.

Enlarge the screenshot

the numbers: 6 points behind Sunburst, ahead of Flare

On the Artificial Analysis editing leaderboard published with the announcement, MAI-Image 2.7 scores 1,193 Elo, against 1,199 for GPT Image 2.5 Sunburst at max quality and 1,182 for GPT Image 2.5 Flare at max quality.

Microsoft's chart ranks eight models from Artificial Analysis blind votes. GPT Image 2.5 Sunburst (max) leads with 1,199. MAI-Image 2.7 sits right behind at 1,193. Then comes GPT Image 2.5 Flare (max) at 1,182, and a three-way tie at exactly 1,151: MAI-Image 2.6, Krea 3 (Ultra) and Nano Banana 2.1. MAI-Image 2.6 Flash follows at 1,144 and GPT Image 2 (high) closes the list at 1,136.

That is a 42-point jump over version 2.6 in a few weeks. For Arena, Microsoft AI posted on X the same day: a 45-point gain over 2.6, third place on the image edit leaderboard, second place by lab, ahead of GPT Image 2.

Still, six Elo points is a very small gap. These scores are voter preferences between two images, not a strict quality measurement, and two models this close can swap places from one week to the next. The chart is Microsoft's own, published for its own launch, and OpenAI's models are measured at their most expensive setting. So here is what I would take from it: MAI-Image 2.7 now plays in the same league as Sunburst, and Microsoft is no longer an outsider in image editing.

Enlarge the screenshot

where it really improved: reasoning, style and text

According to Microsoft, MAI-Image 2.7 reaches first place in four Artificial Analysis categories, with its biggest jumps in reasoning-based edits and scene and style (from 7th to 1st), then text and symbols (from 6th to 1st).

The second chart says more than the overall ranking. It compares where 2.6 and 2.7 place in each category, in an evaluation dated October 22, 2026. Reasoning-based edits: from 7th to 1st. Scene and style: from 7th to 1st. Text and symbols: from 6th to 1st. Microsoft adds enhancement and restoration to the categories where the model reaches the top, and claims a 24-point lead over the next model in scene and style.

What does that mean in practice? A reasoning-based edit is an instruction where the model has to understand the scene before it touches a pixel: put a glass on the table and get its reflection too, switch off a lamp and have the room darken with it. Scene and style means taking a photo shot under office fluorescents and moving it to late-afternoon light without changing the framing. And text has been the weak spot of every one of these models: changing the name on a shop sign while keeping its perspective, its typeface and the wear on the board.

That last one is where I want to see the first real tests. Generation models have written clean text for a year now, as our comparison of the best AI image generators for text shows. Replacing text that is already in a photo without damaging the rest of it is a much harder job.

Enlarge the screenshot

what we still don't know

No API, no price, no date for text-to-image, and as of November 1, 2026 the Vercel AI Gateway catalog only offers MAI-Image 2.6 and 2.6 Flash.

On the terms side, Microsoft has published no pricing, no usage limits, no API and no commercial terms specific to this model. One small detail, spotted by OpenTools: the MAI Playground link on the announcement page still has "mai-image-2-6" in its address. If you test there, check the model name on screen before drawing conclusions.

Arena and Artificial Analysis are built for comparing models, not for production work. You can form an opinion from a handful of edits, but you won't build a workflow on them. And until text-to-image ships, MAI-Image 2.7 replaces no creation model: you have to bring it a starting image.

On November 1 we checked the Vercel AI Gateway catalog, which most Imaginode models go through. On the image side, Microsoft offers MAI-Image 2.6 and MAI-Image 2.6 Flash there, not 2.7. If it shows up, it will go the same way as every other model: a real test request, a measured cost, then a credit price shown before you generate.

meanwhile, MAI-Image 2.6 already edits in Imaginode

MAI-Image 2.6 and 2.6 Flash have been in Imaginode's Image node since October 15, 2026, at 12 and 5 credits per image, with up to five input images for an edit.

The previous version is already usable without a Microsoft account. In Imaginode, add an Image node, pick MAI-Image 2.6 or its Flash version, connect up to five input images and describe the change in one sentence: the bottle from the first photo placed in the scene of the second, a scarf added to a portrait. The price shows on the button before you run it.

Microsoft's own leaderboard leaves little room for doubt. At 1,151, version 2.6 ties with Nano Banana 2.1 and sits 48 points behind GPT Image 2.5 Sunburst, which costs 3 credits per image on Imaginode at Standard and 9 at Max. For a precise edit today, Sunburst is the one I would pick, and our GPT Image 2.5 article shows three of its edits. Version 2.6 is still handy for comparing: two nodes side by side, the same input image, the same instruction, and within a minute you see which one respects your photo better. The Flash version, at 5 credits, works as a cheap draft before you move to a pricier model.

Enlarge the screenshot

what it changes in the way you edit

The race has moved from generation to editing, but the method stays the same: one approved image, one change per instruction, and the previous version kept next to the new one.

On September 8, OpenAI released Sunburst, billed as its most precise editing model, and Microsoft now answers with a model that does nothing else. Both are saying the same thing: a good-looking image is no longer enough, you need to be able to fix it without breaking it. Good news for anyone who edits product shots or client visuals all day.

The model won't do all the work, though. An instruction that asks for three changes at once drifts on any model. An instruction that starts with "change only" and names what must stay put holds up much better. In a canvas, the approved image stays in its node, the edit goes into the next node, and you compare both versions side by side instead of scrolling back through a chat. Our guide on how to recreate the same image covers these habits.

To practice on the categories where MAI-Image 2.7 improved, start with three tools: change a photo background, restore an old photo and the lighting test on a single scene. When 2.7 leaves preview, you will know exactly what to ask it.