Microsoft previews MAI Image 2.7 editor, ranked No. 2 in human votes

Microsoft released MAI Image 2.7 on Oct. 27 as a preview model for editing existing images, not for making new ones from a text prompt. On an image-editing leaderboard built from human votes, it ranks No. 2 with a score of 1,193 Elo. Microsoft says its text-to-image mode is coming soon but has given no release date.

The launch fits the direction of the field. Image models are now judged less on a first render and more on how cleanly they change one part of a picture and leave the rest alone. The preview also leaves gaps that builders will notice: no listed API, no pricing and no general availability date.

What MAI Image 2.7 Preview does

The preview covers three kinds of work on an existing image, according to TokenPost’s report:

• Restyling and relighting: changing the look or the lighting of a scene.
• Text and symbol edits: changing lettering and symbols that appear inside the image.
• Restoration and enhancement: repairing or improving existing images.

Editing text inside a picture continues a theme from earlier in the MAI line. When Microsoft announced MAI Image 2 in March, it singled out clearer text within visuals as one of the main upgrades. It said the model “enables consistent creation of infographics, slides, diagrams, and more”.

Item	MAI Image 2.7 Preview
Introduced	Oct. 27, 2026
Status	Preview
Mode	Editing existing images; text-to-image “coming soon”, no date
Editing tasks	Restyle, relight, change in-image text and symbols, restore and enhance
Leaderboard position	No. 2 on an image-editing leaderboard
Score	1,193 Elo across 4,891 samples
95% Elo interval	1,178 to 1,208
API	Not listed as available on the leaderboard
Pricing, general availability	Not disclosed

What changed from MAI Image 2.6, by Microsoft’s account

Microsoft says MAI Image 2.7 improves on MAI Image 2.6 across all 10 use cases and seven editing actions that it tested. The company also reports a 24-Elo lead over the next model on scene and style tasks. Both figures are Microsoft’s own evaluation. The available coverage does not name the comparison models, the test prompts or how the use cases were defined. Treat them as the maker’s claims until an independent breakdown appears.

The predecessor is recent. Coverage of MAI Image 2.6 cites a September 2026 Microsoft launch note. That note says both the flagship model and a faster “Flash” variant support multi-reference editing (combining several input images) and dynamic aspect ratios. Atlas Cloud’s guide describes a workflow where MAI Image 2.6 Flash Text-to-Image makes a source scene and a separate MAI Image 2.6 Edit model handles the controlled change. The 2.7 coverage does not say whether 2.7 keeps multi-reference input, which aspect ratios it supports, or whether a Flash tier will follow.

How to read the No. 2 ranking

The leaderboard places MAI Image 2.7 second, with 1,193 Elo across 4,891 samples. Elo is a rating system borrowed from chess. Here, people see two edits side by side and pick the one they prefer. A model’s score rises or falls with each win or loss, and the samples are the individual votes behind the rating.

The key caveat is the confidence interval, the range in which the true score most likely sits. MAI Image 2.7’s 95% interval runs from 1,178 to 1,208. TokenPost notes that this overlaps the first-ranked model’s interval, so the placement alone does not establish a clear performance gap. In practice, the top two are statistically close on this board. Whether 2.7 is better or worse than the leader for a given job is not settled by the ranking.

Three more limits apply:

• The first-place model is not named in the available report, and neither is the margin between the two point scores.
• Rankings differ between leaderboards. A separate composite index from HuaLinYuan, which uses its own testing rather than pairwise votes, lists OpenAI’s GPT Image 2.5 Sunburst first at 1524. It places Microsoft’s earlier MAI-Image 2.6 fifth at 1428, tied in score with Google’s Nano Banana 2.1. That index uses a different scale and does not yet show 2.7, so its numbers cannot be compared directly with the 1,193 figure.
• Preference is not preservation. A vote records which result a person liked. It does not directly measure whether untouched parts of the image survived. Edit-model reviewers increasingly treat preservation as the real test.

Where MAI Image 2.7 sits among rival editing models

Editing models have arrived quickly this autumn, and most of them already offer a working endpoint. According to each::labs, OpenAI released GPT Image 2.5 on Sept. 8, 2026, as two models. Flare is the faster one and Sunburst the quality-focused one, and both have edit endpoints that accept up to 16 source images and an optional mask. A mask is an overlay that marks the area to change. Google made Nano Banana 2.1 generally available on Oct. 6, 2026; its edit endpoint takes 1 to 14 images per request. Krea introduced its Krea 3 editing model on Oct. 15, 2026, and says most edits finish in under 8 seconds at native 2K, by its own measurement. Black Forest Labs’ Flux 3 Edit targets regions with bounding boxes instead of masks.

Against that field, the 2.7 coverage gives quality claims and a vote-based ranking. It gives none of the operational details builders usually compare: input image limits, mask or box support, output resolution, latency or price. Microsoft’s earlier MAI models did reach developers through Microsoft Foundry. The GenAI PM wiki describes MAI-Image-2 launching on the MAI Playground and Microsoft Foundry, with a rollout planned for Copilot and Bing Image Creator. Digit reported in March that MAI Image 2 API access was available to select customers such as WPP. Nothing in the 2.7 report says whether the new model will follow the same route.

What it means for builders

For now, MAI Image 2.7 is something to track, not something to integrate. The leaderboard does not list an API, and Microsoft has not disclosed pricing or general availability. Teams can still plan around a few points.

The task list maps to common production jobs

Relighting, restyling, fixing in-image text and restoring old or damaged images are everyday tasks in marketing, e-commerce and archive work. Editing text and symbols inside an image matters for packaging, signage and localised creative, where a wrong letter makes an asset unusable.

Test for drift, not just for the best single result

Reviewers of edit models note a common failure: a requested change quietly alters parts of the image that were never mentioned. One example is a label that starts to “melt” when only the light was meant to change. The same concern runs through Atlas Cloud’s MAI Image 2.6 guide. It advises making one isolated change at a time, naming every detail that must survive, and comparing the result against the source at the same size. Those checks are a sensible starting point for evaluating 2.7 once it is accessible. They also test something a preference vote does not capture.

Expect a split model line-up

Microsoft has shipped its image models in tiers before. GenAI PM describes MAI-Image-2-Efficient as a high-volume, cost-controlled option, set against MAI-Image-2 as the higher-fidelity choice. MAI Image 2.6 came with a Flash variant alongside the flagship. With 2.7, editing arrives first and text-to-image follows at an unannounced date. Pipelines that need both modes may have to call separate models or wait.

Open questions

• Which model leads, and by how much? The report confirms the overlapping intervals but does not identify the first-ranked model.
• What did Microsoft test? The 10 use cases, seven editing actions and the 24-Elo lead on scene and style tasks come without a published method in the available coverage.
• When does text-to-image ship? Microsoft says “coming soon” with no date.
• How will developers get access? There is no listed API, no pricing and no general availability information.
• What are the input and output limits? The coverage does not state resolution, aspect ratios, reference-image support or a licence.

What to watch next

The next concrete signals are an API listing and pricing from Microsoft, and a date for the text-to-image mode it calls coming soon. As more votes arrive on the leaderboard, 2.7’s interval should narrow. That will show whether its No. 2 position separates from the leader or from the models below it; with 4,891 samples so far, the range still overlaps first place. Builders should also watch for a technical write-up that explains Microsoft’s 10-use-case comparison with MAI Image 2.6. The rivals’ edit endpoints are already live, so independent side-by-side preservation tests will matter once 2.7 can be called directly.