MAI-Image-2.7-Preview vs MAI-Image-2.6-Preview: +42 Elo, Minus a Mode

A 42-point Elo gain usually earns a migration. Here it earns a test. MAI-Image-2.7-Preview is a clearly better editor than MAI-Image-2.6-Preview — the gap is wide enough to change which edits you trust the model with — but it is not a drop-in replacement, because the newer model does not generate images from text at all. Microsoft shipped 2.7 on October 27, 2026 as an editing-only preview and put text-to-image on a "coming soon" list. Anyone with a pipeline that calls one model to both generate and edit is therefore looking at an upgrade that covers half their workload.

That makes this comparison less about which number is larger and more about which model fits which job. The two models share a family, a research team and a scoring board. They do not share a feature set, an access path, or a price.

Same ruler, 42 points apart

On Artificial Analysis's AA-Image-Editing v2.1 board, evaluated October 22, 2026, MAI-Image-2.7-Preview scores 1,193 Elo and sits second overall. MAI-Image-2.6-Preview scores 1,151 and sits in a three-way tie with Krea 3 (Ultra) and Nano Banana 2.1; its faster sibling, MAI-Image-2.6-Flash, scores 1,144. The 42-point gap between the two Microsoft models is the number Microsoft leads with, and on this board it is a clean, like-for-like comparison: same evaluation version, same voters, same snapshot.

Two context notes keep that number honest. First, these figures come from Microsoft's launch post reproducing the Artificial Analysis board; the board is independent, the reproduction is not, and we have not re-run it. Second, do not compare either score to anything published before the v2.1 rebuild. Microsoft's September release put MAI-Image-2.6 at No. 1 on the editing board of that era, and it would be easy — and wrong — to read the current 1,151 as a regression. The board changed. Within v2.1, 2.6 is a 1,151 model.

The three categories that flipped

Microsoft says 2.7 improves on 2.6 across all 10 use cases and all 7 editing actions Artificial Analysis evaluates. The headline is the category ranking: reasoning-based edits moved from No. 7 to No. 1, scene and style from No. 7 to No. 1, and text or symbol edits from No. 6 to No. 1. Microsoft also claims the category frontier in enhancement and restoration, and says its widest lead is scene and style, 24 Elo ahead of the next-best model.

For a 2.6 user, the rank flips are more informative than the overall gap. A model that was seventh at reasoning-based edits was, in practice, a model you did not trust with an instruction like "move the product to a marble counter and match the window light." A model that is first is one you can put in front of that job. The same applies to text and symbol replacement, which is the least forgiving edit in commercial imagery: a label that warps or a logo that changes weight is an instant reject, regardless of how good the rest of the frame looks.

The usual caveat applies: category scores are cuts of the same preference data, not independent experiments. But the direction is consistent with the overall score, and it is consistent with the examples Microsoft chose to show — restyling and relighting whole scenes, changing text and symbols in place, restoring existing assets. Those are the edits where a generation-first model tends to struggle, and where an editing-first model can specialize.

The missing mode is the whole ballgame for some pipelines

MAI-Image-2.6-Preview is a unified model: text-to-image and image-to-image edits through the same deployment, with multi-image reference editing, web grounding and model-chosen aspect ratios. MAI-Image-2.7-Preview does edits. Microsoft's post is explicit that generation is "coming soon," and the model's own launch examples are all transformations of an existing image.

That single difference decides the comparison for a large class of users. If your workflow starts from a prompt — a campaign concept, a product visual, a blog hero — 2.7 has nothing to offer you today. If your workflow starts from an asset — a client's photo, a scan, a packaging mock — 2.7 is the better tool and 2.6's generation capability is irrelevant. The version number implies a superset. The product is a specialization.

It also changes what "upgrade" means for a team. Moving from 2.6 to 2.7 is not a model swap; it is a split. Generation stays on 2.6, editing moves to 2.7, and the two paths now have different availability, different limits and, eventually, different prices. That is a more expensive architecture than one model behind one endpoint, and it is worth doing only if the editing gain is large enough to pay for the complexity. For scene restyling and text replacement, it might be.

Access and cost: one is buyable, one is not

MAI-Image-2.6-Preview is available in Microsoft Foundry as a public preview, alongside MAI-Image-2.6-Flash. Its published rates are $5 per million text input tokens, $8 per million image input tokens and $38 per million image output tokens; Flash is $1.75, $2.50 and $19. Artificial Analysis estimated those at roughly four cents per image for 2.6 and about two cents for Flash. The documented limits are five reference images and a maximum output of 2,359,296 pixels — 1536×1536 — with 768-pixel minimums on each dimension, plus web grounding and automatic aspect-ratio selection.

MAI-Image-2.7-Preview has none of that. It is in MAI Playground, on Arena and on Artificial Analysis. There is no Foundry entry, no version string to deploy, no rate card, and no published reference-image or resolution limit. Microsoft has not said whether 2.7 will inherit 2.6's five-reference ceiling, exceed it, or ship with different controls. In a comparison about capability, that is an absence of evidence, not evidence of a downgrade — but it is also not something a production team can plan around.

• Editing score: 2.7 at 1,193 vs 2.6 at 1,151 (+42)

• Generation: 2.7 not available at launch vs 2.6 text-to-image and editing

• Access: 2.7 in MAI Playground, Arena, Artificial Analysis vs 2.6 in Microsoft Foundry

• Price: 2.7 unpublished vs 2.6 at $5 / $8 / $38 per million tokens, about $0.039 per image

• Limits: 2.7 unpublished vs 2.6 at five references and 1536×1536 output

• Fast variant: none for 2.7 vs 2.6-Flash at $1.75 / $2.50 / $19, about $0.020 per image

What 2.6 still does better

Beyond generation, 2.6 owns every operational advantage. It has a documented API surface, a deployment version, a rate card, a Flash sibling for throughput, and a Foundry contract that an enterprise can put through procurement. It also has features 2.7 has not been described as carrying: web grounding, automatic aspect-ratio selection, and a published multi-image reference workflow.

Some of those may simply be undocumented for 2.7 rather than absent. Microsoft's launch post is short and focused on editing quality; it does not read like a full model card. But the burden of proof runs the other way for a production decision: a feature you cannot find in the documentation is a feature you cannot depend on.

The Flash comparison is the one to watch for volume work. MAI-Image-2.6-Flash is not the quality leader — 1,144 Elo, seven points behind the main 2.6 — but at roughly two cents per image it is in a different cost class from anything 2.7 is likely to launch at. If Microsoft prices 2.7 like a flagship, the rational stack for many teams will be Flash for bulk edits and 2.7 for the hard cases, with the main 2.6 squeezed out entirely.

If you already build on 2.6

Do not migrate the generation path. There is nothing to migrate it to. Keep 2.6 or 2.6-Flash for text-to-image and for any edit that is routine enough that a seven-Elo difference does not matter.

Do run a shadow test of 2.7 on the edits that currently fail. Pull the twenty assets your team has had to redo by hand — the label that would not sit straight, the product that would not relight, the scan that lost detail — and run them through the playground. If the category claims hold on your own material, 2.7 earns a place as the escalation path: 2.6-Flash for volume, 2.6 for standard work, 2.7 for the assets that need a human pass today.

Neither MAI model is on OrcaRouter, so this is not a routing decision we can make for you; Microsoft's own surfaces are the only way in. If what you actually need is a callable editing endpoint with failover behind one key, the routed alternative is GPT-Image-2.5 — Sunburst for quality, Flare for throughput — at 0% markup, provider list price passed through. That is a different model family with a different look, not a substitute for 2.7's scene-and-style lead, and it should be evaluated as such.

What would make 2.7 the default

Three things. Text-to-image parity, which would turn a specialist into a replacement. A Foundry listing with a price at or below 2.6's, which would make the editing gain affordable at volume. And published operational limits — references, resolution, throughput — which would let teams plan instead of guess. Microsoft delivered all three for 2.6 in a single September release, so none of this is speculative capability work; it is packaging.

Until then, the right reading of the 42 Elo is narrow and useful: Microsoft's editing model got materially better, and its generation model did not change. The version number went up. The product line split. Plan for the split, not the bump.