Microsoft's MAI Image 2.7 edits only the parts you pick and narrowly beats OpenAI

At Ignite, Microsoft unveiled its MAI Image 2.7 image model along with a faster Flash variant. Both are built to change individual objects in a photo and leave the rest of the image alone. On an independent image-editing leaderboard, the model sits just ahead of OpenAI and costs a fraction of the price.

Model releases Nov 18, 2026 6 min read Source: Bing News (nl-NL)
Illustrative image, AI-generated

What's going on

At its Ignite developer conference, Microsoft introduced two new image models: MAI Image 2.7 and a faster, cheaper version called MAI Image 2.7 Flash. Both come from Microsoft AI (MAI), the company's in-house AI division. The focus is on image editing: you ask the AI to change one specific part of a photo, and everything else should stay the same. Microsoft is also launching Canvas, an image editor in MAI Labs, its sandbox for new AI features. Both models are in public preview and available through the Microsoft Foundry developer platform and through partners.

According to the trade publication The Decoder, the image models are part of a bigger announcement. Microsoft released seven specialized models at once at Ignite, including models for coding (MAI Code 2), cybersecurity (MAI Cyber 2 Flash) and speech (MAI Realtime 1). The pitch is the same across the board: nearly as good as OpenAI and Anthropic, but much cheaper.

What the model can do

Microsoft says MAI Image 2.7 can selectively change the color, material or position of individual objects. The Dutch site aiwereld.nl gives the example of a chair in a room: it gets new upholstery while the rest of the room stays exactly as it was. According to Microsoft, the model also holds on to details better when you ask for several edits in a row. How well that actually works depends on the photo and the instruction, aiwereld notes.

Both versions can also generate images with transparent backgrounds. That's handy for catalog product shots you later drop onto a white or colored background. You can also generate several versions of an object without the AI rebuilding the whole scene each time.

In the Canvas editor, which runs on the Flash variant, you start with a text prompt or a sketch. Then you mark an area of the image and have it moved, replaced or redrawn in a different style. Microsoft also suggests using it to turn a rough sketch into a diagram. Flash is aimed at anyone who wants to see lots of variations quickly.

aiwereld recommends some caution. Shadows, small text and object sizes can shift unintentionally, even when you're asking the model to edit something else entirely. Putting two catalog photos side by side makes these discrepancies easier to spot.

Benchmarks: narrowly ahead, but within the margin of error

The ranking comes from the Artificial Analysis leaderboard. There, people see two anonymously edited images side by side and pick the better one. The platform turns these votes into an Elo score, a rating system borrowed from chess: beat strong opponents and you climb faster.

Model Elo (image editing) Price per 1,000 images*
MAI Image 2.7 1,215 approx. $35
GPT Image 2.6 Sunburst (max) 1,211 approx. $211
MAI Image 2.7 Flash – approx. $10
Nano Banana 2.1 (Google) – approx. $34
Nano Banana Pro (Google) – approx. $134

*Prices according to Microsoft (Foundry) or the Artificial Analysis comparison, as cited by Decrypt. Your actual bill depends on provider, resolution and settings.

According to The Decoder, the 1,215 score is based on 5,478 votes. Decrypt points out that with a gap of just four points, the confidence intervals of the two top models overlap. Statistically, it's closer to a tie. In pure text-to-image generation, Microsoft only ranks fourth on the same leaderboard.

Crypto Briefing had already covered an earlier preview, which it said came out on October 27. Back then, MAI Image 2.7 ranked second with 1,193 Elo, behind OpenAI's GPT Image 2.5 Sunburst at 1,199. On the Arena Edit leaderboard it came third. According to Crypto Briefing, the model gained 42 points over its predecessor, MAI Image 2.6 from August. It improved in all ten use cases and all seven edit types that Artificial Analysis measures. The biggest jumps came in composition and cropping (+57 Elo) and in-image text edits (+49 Elo). Pricing hadn't been announced at the time. That changed with the Ignite announcement.

First outside impressions

Decrypt ran a brief hands-on test. In a handful of head-to-head comparisons, its images beat GPT Image and Google's Nano Banana on realism, creativity and overall quality. The outlet stresses that these are only first impressions. Earlier tests of MAI Image 2 had shown excellent realism, Decrypt says, but also frustrating limits in editing and output formats. Canvas, transparent backgrounds and Foundry access target exactly those weaknesses. Still, Decrypt notes that great image quality and a great product are two different things. It also found the model fairly heavily censored.

Where you can use it

If you want to try the new models, you'll find them in MAI Labs (Canvas), in Microsoft Foundry and with partners. According to The Decoder, Adobe, Figma Weave, Minute (Sports Illustrated), Wirestock and Manus work with MAI Image. Since June, a model from the MAI Image series has been the default behind Bing Image Creator. According to aiwereld, though, that refers to the series in general, not specifically to the new versions. Decrypt reports that select enterprise customers can fine-tune and run licensed model weights in Azure. That doesn't make the technology open source, though.

The bigger picture

Microsoft is moving fast to reduce its reliance on outside providers, above all its own partner OpenAI. According to Crypto Briefing, MAI Image 2.7 is already the third image model in about seven months, after MAI Image 2 in March and 2.6 in August. Price is its strongest selling point. For practically the same editing quality, Microsoft charges only about a sixth of what GPT Image 2.6 Sunburst costs at maximum quality. For companies editing thousands of product images, that can add up fast.

There are caveats, though. The leaderboard lead is within the margin of error. The model doesn't lead in text-to-image generation. And it has no price advantage over Google's Nano Banana 2.1. Microsoft's claims about preserving detail across multiple edits come from Microsoft itself and haven't been independently verified yet. If you need targeted photo editing, this is a serious, affordable alternative. Whether it holds up better in everyday use will take longer testing to tell.

Related videos

Microsoft Ignite: Opening Keynote 2026 Microsoft · YouTube

Sources

#microsoft#image editing#mai image#ignite#image generation
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.