MAI Image 2.7 Preview Leads Four of Seven Edit Actions — and Cannot Generate a Picture

Microsoft AI released MAI Image 2.7 Preview on October 27, 2026, and the most useful thing to know about it is a sentence in the announcement itself: the model starts with editing, with text-to-image generation "coming soon." Everything that follows — the second-place finish on Artificial Analysis, the third-place finish on Arena, the 42-point gain over MAI-Image-2.6 — describes a model that can improve an image you already have and cannot make one from nothing.

That is a narrower product than a leaderboard position implies. A generative image model is a blank-page tool; an editing model is a repair and revision tool. Most of the jobs people actually pay for in 2026 are revision jobs, which is why the editing board has become the one that commercial buyers watch. But the split matters when you sit down to try it: you will need a source image before MAI Image 2.7 Preview has anything to do.

The model is available in MAI Playground, on Arena, and on Artificial Analysis. There is no API, no Microsoft Foundry listing, and no published price. Microsoft has not disclosed 2.7-specific limits on reference images, resolution, or output tiers.

What "Editing" Covers Here

Microsoft's description of the model is unusually specific about the edits it was built for: whole-scene restyling, relighting, changing text and symbols while keeping them integrated into the image, restoring or enhancing an existing asset, and edits that require understanding objects, lighting and environments together. That list is not generic. Each item names a failure mode that earlier editing models handled badly — the relight that leaves shadows pointing the wrong way, the new label text that floats above the surface instead of sitting on it, the upscale that sharpens noise along with detail.

The integration phrase is the one worth reading twice. Text editing in image models is a solved problem in isolation and an unsolved one in context: any model can render a word, and most of them render it on top of the scene rather than into it, with the wrong perspective, the wrong motion blur, and a font that does not match the surface it lands on. Microsoft is claiming that specific gap as a strength, which is a checkable claim rather than a mood.

Four of Seven, and the Three That Went the Other Way

Artificial Analysis scored MAI Image 2.7 Preview at 1,193 Elo on AA-Image-Editing v2.1, using a snapshot dated October 22, 2026. That puts it second, six points behind GPT Image 2.5 Sunburst at its max quality tier (1,199) and eleven ahead of GPT Image 2.5 Flare (1,182). The model it replaces, MAI-Image-2.6, sits at 1,151 — a 42-point improvement.

The category breakdown is more informative than the aggregate. According to Artificial Analysis, 2.7 ranks first on four of the seven editing actions it evaluates, and Microsoft says the model improved across all ten use cases and all seven actions. The four it leads are enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing. The widest margin is scene and style, where Microsoft claims a 24-point lead over the next-best model.

The three it does not lead are composition, individual-object edits, and identity preservation — all three held by OpenAI. That is a coherent picture rather than a random scatter. OpenAI's models are strongest exactly where the instruction has to be followed literally and locally: move this object, keep this face, respect this layout. Microsoft's model is strongest where the instruction requires judgment across the whole frame: relight the scene, restyle it, restore it, rewrite the sign.

The Text-and-Symbol Result Is the Commercially Interesting One

Of the four categories 2.7 leads, text and symbol editing is the one with the clearest business case outside a creative studio. Packaging mockups, retail signage, slide decks with chart labels baked into an exported PNG, product shots with a regional variant of the same label — all of these are edits where the expensive part is not drawing the new word but preserving the lighting, perspective and material response of the surface it sits on.

It is also the category where a preview is least useful and most frustrating. A designer who wants to fix a label inside a photograph can do it in Playground, one image at a time, by hand. A pipeline that fixes five thousand labels cannot, because there is no endpoint to call. The capability and the delivery mechanism are mismatched, and that mismatch is the actual news of this launch — not the Elo number.

Scene and style is the opposite case. A 24-point lead in whole-frame restyling is the kind of margin that shows up in blind preference tests, and restyling is a batch job by nature: one source frame, forty looks. A model that wins that category by a wide margin is a model a studio would want in a loop, and there is no loop available.

What "Preview" Has Meant at Microsoft Before

The predecessor is the best available guide to the timeline. MAI-Image-2.6 entered Microsoft Foundry on September 4, 2026, roughly six weeks before 2.7 appeared. Its published rate card is per token rather than per image — $5 per million text input tokens, $8 per million image input tokens, $38 per million image output tokens — which works out to roughly four cents for a typical image. The Flash variant runs $1.75, $2.50 and $19, or about two cents an image.

If 2.7 follows the same path, the missing pieces are a Foundry listing, a version string, and a rate card. The distance between "second on the editing board" and "a line item in your budget" is exactly those three artifacts. Until they exist, a team that needs an editing model in production today is choosing between 2.6 in Foundry and whatever else it can already call.

How to Test It This Week

MAI Playground is at playground.microsoft.ai and takes a Microsoft account. Arena exposes the model for blind side-by-side voting against other editors, which is a genuinely useful way to check the scene-and-style claim yourself: put a photograph in, ask for a relight, and see which output you prefer when you do not know which model produced it.

Two caveats before you plan around it. First, Microsoft has published no 2.7-specific generation caps. The free-tier limits that circulated during the MAI-Image-2 era — on the order of fifteen generations a day with a wait between them — were never confirmed for this model and should not be assumed. Second, an Arena vote is not a product trial. Arena tells you which output people prefer under a neutral prompt; it does not tell you how the model behaves on the fortieth revision of the same asset, which is the actual working condition.

What It Would Take to Move This From Playground to Production

Three things, in order of importance. A deployment target — Foundry or a public API — because without one the model is a demo. A price, because "second on the editing board" is not a purchasing argument until you can divide it by cost per image. And text-to-image parity, because a great many real editing workflows begin by generating the base image before revising it, and a model that cannot start a picture forces every pipeline to bolt on a second vendor.

For teams that need an editing model under one key now, the practical answer is different from the interesting answer. OrcaRouter does not host MAI Image 2.7 Preview — no one does, because there is no API — but GPT Image 2.5 Sunburst and GPT Image 2.5 Flare, the two models directly above and below it on that board, are both available through the gateway at their provider list price with no markup added. That makes the six-point gap between first and second place testable today in a way the second-place model is not: you can run the same edit through both OpenAI tiers and see which one your own assets prefer, then keep the winner in the routing config.

The honest summary is that Microsoft shipped the more interesting editor and OpenAI still ships the more available one. The six Elo separating them is smaller than the distance between a playground and an endpoint, and only one of those gaps closes on its own.