Changing the light in a photograph sounds simple to ask for. Getting that lighting without moving an object, distorting a face or altering an inscription is another matter. With MAI Image 2.7, Microsoft is putting the emphasis squarely on controlling changes made to an existing image.
The announcement on 27 October 2026 describes a preview intended for editing: the style and lighting of a scene, changes to text or symbols, restoration and image enhancement. Microsoft says the model can be tried in MAI Playground, Arena and Artificial Analysis. Generation from a single text prompt is announced for a later stage, and is not presented as available in this first phase.
The new release should therefore be read as an image-editing proposition. It is not enough to establish that one and the same system would be the best for every kind of visual creation.
The ranking: a specific point of reference, not a universal verdict
In the AA-Image-Editing v2.1 table consulted on 4 November 2026, MAI-Image-2.7-Preview ranks second, with an Elo score of 1,196. GPT Image 2.5 Sunburst (max) is ahead at 1,199, and GPT Image 2.5 Flare (max) is behind at 1,183.
The detail matters: simply claiming that “Microsoft beats OpenAI” obscures the existence of several competing versions. Artificial Analysis ranks models here on editing an image in response to an instruction, according to human preferences between two results.
The table also gives 95% intervals: from 1,181 to 1,211 for MAI Image 2.7 and from 1,188 to 1,210 for Sunburst. Their overlap is a reason not to treat the three-point difference as a clear-cut gap. These figures are a snapshot of the ranking, which may change.
This type of evaluation helps identify tools worth examining. It does not replace testing against the specific requirements of a newsroom, a studio or a retailer.
A good edit preserves what was meant to stay
An editing instruction often has two parts: what should change and what should remain the same. The second part is sometimes implicit. Someone asking for a warmer atmosphere is not necessarily asking for new architecture, different clothing or a change to the text on a box.
A result can therefore be appealing at first glance and still be unusable. A more elegant shadow does not erase the fact that a product reference has become wrong. A better-balanced composition does not make up for the disappearance of a detail the client wanted to keep.
This perspective changes how a tool should be evaluated. Instead of asking only whether the final image looks good, we need to check whether it fulfils the precise request and respects the elements that were meant to be protected. Editing is judged as much by what it refrains from doing as by what it changes.
A fidelity test before a series of variations
Imagine a product photograph bearing a label. The aim would be to lighten the background without touching the product’s shape, colour or text. This scenario is an example of a test protocol; we have not run it on MAI Image 2.7.
The first comparison would focus on the requested change: is the background actually lighter? The second would examine the elements to be preserved: the same proportions, the same inscription, the same distinctive details. The third would look at the joins, shadows and areas close to the outline.
To avoid choosing only the prettiest attempt, the number of attempts and the corrections required would need to be recorded. A result achieved once after many rounds of revision does not tell us the same thing as a result that can be reproduced across several comparable images.
This test can remain modest. A few well-chosen images, with clear criteria, sometimes provide more useful information than a long gallery where nobody knows what was originally requested.
Restoring a detail is not recovering evidence
The term “restoration” calls for particular caution. Making an old image clearer or more pleasing does not prove that every newly visible detail corresponds to what was in front of the camera.
For an artistic creation, an interpretation can be acknowledged as such. For a documentary archive, a product image or a photograph used to establish a fact, the stakes are different. It must be possible to distinguish the original from the changes and avoid presenting a reconstruction as an observation.
This question concerns the use of generative tools in general; it is not a flaw measured here in MAI Image 2.7. It is a reminder that the quality of an image also depends on what we claim to show the reader.
Preview, access and cost: three separate checks
Microsoft presents its model as a preview. At the time of our consultation, Artificial Analysis indicated that no API was available for this leaderboard entry. Being able to try an interface therefore does not guarantee that the same model can be integrated into an automated workflow.
For professional use, the actual access conditions, the version used, the formats accepted and the rules that apply to submitted images still need to be checked. No MAI Image 2.7 API price is inferred here from that of an earlier version.
The effective cost would not, moreover, be limited to a single attempt. It would include rejected variations, checking and additional editing until an image was accepted. That final result would make it possible to compare tools in operational terms.
MAI Image 2.7 thus offers an interesting avenue for image editing, backed by a high ranking and targeted functions that have been announced. To decide whether it is useful in real work, the most productive question remains a concrete one: can it make the requested change while leaving intact what we wanted to keep?
Illustrative photograph: the Microsoft Visitor Center in Redmond, in May 2022. Credit: Lectrician1 / Wikimedia Commons / CC BY-SA 4.0.
