On October 27, 2026, Microsoft AI announced a preview of its image AI model, MAI Image 2.7. The release starts with image editing, which modifies existing images based on instructions. Text-to-image generation is due to follow soon. On the independent AI evaluation site Artificial Analysis (AA), MAI Image 2.7 scored 1193 Elo in overall image editing, up 42 points from its predecessor MAI Image 2.6, and placed second overall. However, the score alone does not settle how it compares with OpenAI's top models. Anyone weighing a switch from the previous version will need to look at performance for specific editing tasks and at terms that have yet to be announced.
How much can it change while keeping the original image intact?
According to Microsoft, MAI Image 2.7 outperformed 2.6 across all 10 use cases and all 7 editing operations that AA evaluates. The company highlights image quality enhancement and restoration, editing that requires reasoning, and changes to scenes and styles as particular strengths. It also says it improved the editing of text and symbols within images. Reasoning-based editing means making changes that take into account the relationship between objects in an image and their surroundings. Even when only part of an image is altered, how faithfully the model preserves everything outside the instructed area greatly affects the result.
Scene and style changes involve adjusting lighting and the overall look of an image, whereas text editing deals with the wording inside the image itself. So even within image editing, the required abilities differ. Microsoft says that in the scene and style change evaluation, 2.7 beat the next-best model by 24 Elo. Note, however, that the claim of improvement across all 10 use cases and 7 operations is a comparison with the earlier 2.6; it does not mean 2.7 beat competing models in every category. Microsoft also states in its official announcement that the model placed third on Arena Edit, but that is a ranking announced by Microsoft and should be treated separately.
Evaluated almost on par with OpenAI's top models
In AA's image editing benchmark, AA-Image-Editing v2.1, each model receives the same source image and editing instruction, and evaluators compare the results with model names hidden. The Elo score is calculated from votes on which image is preferred; a higher number means the model was more likely to be chosen in those comparisons. Looking at AA's image editing leaderboard as retrieved on October 28, 2026, MAI Image 2.7 is close behind OpenAI's top model. The GPT Image 2.5 Sunburst and GPT Image 2.5 Flare results are both for the max setting.
| Model (setting) | Elo | 95% CI | Samples |
|---|---|---|---|
| GPT Image 2.5 Sunburst (max) | 1199 | 1188–1210 | 18,250 |
| MAI Image 2.7 Preview | 1193 | 1178–1208 | 4,891 |
| GPT Image 2.5 Flare (max) | 1182 | 1171–1193 | 19,136 |
| MAI Image 2.6 | 1151 | 1139–1163 | 16,644 |
| Krea 3 (Ultra) | 1151 | 1137–1165 | 5,505 |
| Nano Banana 2.1 | 1151 | 1138–1164 | 6,868 |
| MAI Image 2.6-Flash | 1144 | 1131–1157 | 10,753 |
Comparing scores at the same point in time, MAI Image 2.7 is 42 Elo above 2.6, just 6 points behind GPT Image 2.5 Sunburst, and 11 points ahead of GPT Image 2.5 Flare. These are simple differences between the listed scores, though, and statistical uncertainty must also be considered. The 95% confidence intervals for 2.7 and the earlier 2.6 do not overlap, while the intervals for Sunburst and Flare do overlap with that of 2.7. A 6-point or 11-point gap is therefore too small to declare a decisive winner. Conversely, overlapping intervals alone do not justify concluding that the models are statistically equivalent.
The number of samples also differs. MAI Image 2.7 has 4,891, compared with 18,250 for Sunburst and 19,136 for Flare, so the two OpenAI models have drawn many more evaluations. That is why it matters to check the confidence intervals, which express the uncertainty of each estimate, alongside the scores. The 42 Elo gain over 2.6 reflects a shift in how often the model is preferred in comparisons; it does not mean a 42% higher editing success rate or faster processing.
The earlier 2.6, Krea 3 (Ultra setting), and Nano Banana 2.1 are all tied at 1151 Elo. But identical overall scores do not guarantee identical strengths in specific kinds of edits. Picking a model simply by working down the leaderboard makes it hard to judge strengths, weaknesses, and fit with an actual production workflow. MAI Image 2.7's No. 2 overall placing is best seen not as grounds for switching on its own, but as a strong data point when comparing models for your own use.
What to check for background, lighting, and text edits
For example, when replacing only the background of a product image, it matters not just whether the background changed as instructed but also whether the product's shape, texture, and color were preserved. When adjusting lighting, you need to check that brightness and mood change without unwanted alterations to parts you did not mention. These are general examples of use cases for assessing the editing abilities Microsoft cites as strengths, not results measured individually for MAI Image 2.7.
Replacing text within an image is also worth testing separately from the overall score. Microsoft lists text and symbol editing as a strength, but whether the model can reproduce specified wording accurately and fix it without changing the surrounding image has to be confirmed with real materials and instructions. In particular, the overall rating alone cannot tell us how accurately it handles Japanese text. Strong results in English text editing do not necessarily carry over to the complex character forms and layouts of Japanese.
When comparing models, it is important to use the same source images and editing instructions. Under identical conditions, it is easier to compare whether the parts meant to stay intact were preserved, whether the intended changes came out as desired, and whether unwanted changes occurred. AA's evaluation also standardizes inputs, but an overall score aggregating the preferences of many evaluators does not necessarily match the requirements a production team has for a particular deliverable. An image that is broadly preferred may still not meet the exact edit that was needed.
In actual production, workflow efficiency changes greatly depending on whether the first output is usable or needs repeated corrections. This ranking does not reveal each model's processing time or how many revisions are needed to reach the desired result. It is therefore not possible to conclude that MAI Image 2.7's quality gains translate directly into shorter production times. Both the ability to make required changes while preserving the original image and the effort needed to reach a finished result need to be verified.
API and pricing are the keys to deciding whether to switch
Microsoft points to three places to try MAI Image 2.7: Arena, AA, and MAI Playground. However, the announcement does not give an API launch date, pricing, or when it will be offered on Microsoft Foundry. AA's API field also reads "No API available," so the conditions for building it into existing production systems are not yet in place.
Availability differs from the 2.6 series. In its announcement on September 4, 2026, Microsoft introduced the quality-focused MAI Image 2.6 and the speed-focused MAI Image 2.6-Flash as a public preview on Foundry. It also mentioned editing with multiple reference images, image generation using web information, variable aspect ratios, and resolution up to 1.5K. These, however, are features and specifications described for the earlier models, and it cannot be assumed that 2.7 offers them under the same conditions.
When comparing prices, the basis for the cost also needs to be consistent. AA's price column lists the cost of generating 1,000 images at 1024×1024 pixels using each developer's API at default settings. GPT Image 2.5 Sunburst and Flare at the max setting are each $210.7, MAI Image 2.6 is $38.9, and 2.6-Flash is $19.5. Nano Banana 2.1 is $33.6, and Krea 3 at the Ultra setting is $270.0.
But these prices are not the total cost of completing one image edit of equivalent quality. They also do not account for the number of revisions or the waiting time needed to reach the desired result, and they do not correspond directly to the editing quality ranking discussed here. Moreover, since MAI Image 2.7's pricing has not been disclosed, the old 2.6 prices cannot be used to call it a cheaper alternative to OpenAI's top models.
Also, Microsoft's September 4 announcement described 2.6 as ranked first on AA's image editing leaderboard, which was based on the evaluation at that time. On the same leaderboard now, 2.6 stands at 1151 Elo and 2.7 at 1193 Elo. Rankings shift as models are added and results are updated, so checking scores and confidence intervals from the same point in time is a better way to gauge improvement than simply comparing rankings from different dates.
MAI Image 2.7 has posted results close to OpenAI's top models in overall image editing evaluation, but there is still not enough information to decide on adopting it in a production environment. Once API terms and pricing are announced, comparisons that include workflow efficiency and cost-effectiveness, not just quality, will become possible. What matters is to use the same materials and instructions and find out how many revisions are needed, and how much time and money it takes, to get the required result. Only that kind of practical testing will show how much a high leaderboard ranking is worth in a real production process.