Second Place on the Editing Board, Absent From the Request Board

On the Artificial Analysis image-editing board, Microsoft's MAI-Image-2.7-Preview sits at 1,193 Elo. Second place, six points behind GPT Image 2.5 Sunburst. On Arena's editing board it is third at 1,471. Across the seven editing actions Artificial Analysis measures, it beats its predecessor on all seven and takes first place in four, including scene and style changes and the reasoning-based category.
On OpenRouter's weekly image-generation request board, it does not appear at all. Not in the top fifty. Meanwhile MAI-Image-2.6 Flash sits at 27th with 35,500 requests, MAI-Image-2.6 at 32nd with 22,200, MAI-Image-2.5 at 33rd, and even MAI-Image-2.5 Pro shows up at 43rd.
One model wins the judge. The other four get the traffic. That gap is the most useful thing I have seen about image models this month.
What "editing only" means in practice
MAI-Image-2.7-Preview does not generate an image from nothing. It takes an image you already have and changes part of it. That is the entire product.
This sounds like a limitation and is actually a thesis. Look at how people use image models in production. An agency is not generating a hero image from a prompt — it generated one last month and the client wants the background warmer, the product bigger in frame, and the model's jacket changed from navy to olive. An e-commerce team has 4,000 product photos shot on white and needs them on marble. A publisher has a layout that is approved and needs the illustration inside it adjusted without touching the type.
All three of those are editing tasks. All three are badly served by a model whose interface is "describe a new image." The difference is not quality. It is whether the rest of the frame stays put.
Microsoft's decision to ship an editing-only model, and to name it a preview, suggests they looked at the leaderboards and noticed something: the generation boards are crowded, and the editing boards have a different competitive structure. On generation, Google and OpenAI have spent two years and enormous compute on a race that is now measured in single-digit Elo. On editing, the top of the board was, until recently, mostly OpenAI plus a cluster of specialists.

The 76% that explains the board
Look at the shape of the OpenRouter data and the editing thesis gets even clearer.
Google owns six of the top ten image models by request volume. Nano Banana 2 alone is 2.19 million requests in a week, 18.5% of a market of 11.81 million. The three OpenAI entries in the top ten total 2.67 million. That is distribution: Gemini is in every Android phone and every Google Workspace account, and GPT Image is the default in ChatGPT.
Now look at the small numbers further down. Krea's K3 Edit is 18th with 114,000 requests, up 751% week over week. That is an editing product, growing faster than anything else on the board. Recraft V4.1 Flash is 12th at 223,000, up 55% — a model with a design-tool lineage, not a chatbot lineage. Flux 3 Image Klein is 15th at 140,000, up 28%, and Flux.2 Klein 4B is 17th at 122,000, up 97%. Those are open-weight models being pulled by a community that wants control.
The pattern: the top of the request board is decided by where the model is embedded. The middle of the board is decided by whether the model does a specific job well. MAI-Image-2.7 is aimed at the middle, and it has not arrived there yet because it is in a preview playground and not in Copilot or Bing.
Why editing is the harder benchmark to trust
There is a reason editing scores feel more meaningful than generation scores, and a reason to be suspicious of both.
Generation benchmarks ask a rater to compare two images against a prompt. That measures aesthetics and prompt adherence, weighted by whatever the rater cares about that day. Editing benchmarks ask whether a specific change was made while everything else was preserved. That is a much more falsifiable question. Either the jacket is olive or it is not. Either the type is intact or it is not.
Which makes the seven-action breakdown the interesting part. Artificial Analysis separates scene and style edits, reasoning-based edits, and five other categories. A model that leads on scene and style but sits fourth on text rendering is a specific tool for a specific job, and that is more useful to a buyer than an aggregate Elo.
The suspicion: 3,402 votes is a small sample. Arena lists MAI-Image-2.7-Preview at 1,471 with a confidence interval of 1,461 to 1,481, on 3,402 votes. GPT-Image-2 has 336,785 votes. When a model is new, its Elo is provisional in a way the leaderboard does not shout about, and a preview with a few thousand votes can look like a podium finish that a million votes would not confirm.
The distribution problem, stated plainly
Here is the trap for anyone reading these boards to make a purchasing decision.
A model can be second-best in the world at editing and have zero production traffic, because the way you get traffic is not by being good. It is by being the default somewhere. Google's advantage on the request board is not that Nano Banana 2 is 60% better than everything else. It is that it is one tap away inside products people already open.
For Microsoft, the path is obvious and slow: put it in Copilot, put it in Designer, put it in the Edge sidebar, put it in Office. The moment that happens, the request board will register it, and the number will not be a measure of quality. It will be a measure of placement.
For everyone else, the lesson is different. If you are choosing a model for a product, the request board tells you what is cheap and available. The Elo boards tell you what is good. Those are two different questions and the answers have diverged sharply.
What to actually do with this
Two concrete things, if you build with image models.
Test on your own edits, not on prompts. Take ten images you have already shipped and ten changes a client actually asked for. Run them through three models. Count how many times the parts you did not want changed stayed unchanged. That number will correlate badly with any leaderboard, and it is the only number your client will notice.
Separate the generate budget from the edit budget. Most teams buy one model and use it for both. The data suggests that is now the wrong shape: generation is a commodity decided by price and placement, editing is where the differentiation still exists and where a 6-point Elo gap can mean a saved afternoon.
MAI-Image-2.7-Preview may or may not become the editing default. But the fact that a major lab shipped a model that cannot generate anything, and put it second on a board, is a signal worth taking seriously. The interesting competition in image models has moved from "what can it make" to "what can it leave alone."
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.