Microsoft AI previewed MAI-Image-2.7 on October 27, 2026, an image model aimed at editing photos, and the numbers circulating with it are that it ranks No. 2 on the Artificial Analysis image-editing leaderboard, behind only OpenAI. Microsoft's own news page now lists a post titled "MAI Image 2.7 debuts above GPT Image 2.5 Flare on Artificial Analysis and #3 on Arena," which already suggests the rank depends on which board and which OpenAI variant you compare against. This post separates what is claimed from what we could confirm, and gives you a way to judge the model without trusting any leaderboard.
A candid note on sourcing: we could read Microsoft AI's news index and its earlier MAI-Image-2.5 launch post, but not the 2.7 post itself or the live leaderboard row. The specifics below come from launch reports that cite Artificial Analysis and Mustafa Suleyman's posts on X, and we mark them as such.
Quick reference: claimed vs verified
| Question | Answer | Status |
|---|---|---|
| What launched? | MAI-Image-2.7-Preview, edit-focused | Listed on Microsoft AI's news page |
| Where can you try it? | MAI Playground, editing only; text-to-image later | Reported, not confirmed by us |
| Rank claimed | No. 2 on AA-Image-Editing v2.1, behind GPT Image 2.5 Sunburst (max) | Attributed to Artificial Analysis and Suleyman |
| Categories led | First in four of seven: scene and style edit, text or symbol edits, reasoning-based edit, enhancement and restoration | Reported from Artificial Analysis posts |
| Where it trails | Composition and framing, object-level edit, identity preserving edit (OpenAI higher) | Reported |
| Another headline | "Above GPT Image 2.5 Flare on Artificial Analysis and #3 on Arena" | Microsoft's own post title, text unread |
| Pricing, API, Foundry | Not found | Unknown |
What was announced
According to the launch reports, MAI-Image-2.7-Preview holds the No. 2 spot on the AA-Image-Editing v2.1 leaderboard, ahead of editing systems from Meta, Google and xAI, and trailing only OpenAI's GPT Image 2.5 Sunburst (max). Artificial Analysis's posts reportedly note that Microsoft's model ranks first in four of seven measured editing categories (scene and style edit, text or symbol edits, reasoning-based edit, and enhancement and restoration), while the OpenAI model still scores higher on composition and framing, object-level edit and identity-preserving edit.
Microsoft AI's chief, Mustafa Suleyman, reportedly posted that MAI-Image is "now the #2 image edit model in the world," beating Google, Meta, xAI and a GPT-Image-2.5 variant. The model is described as a preview available for editing in the MAI Playground, with text-to-image to come.
Two details deserve care. First, the claims come with a category breakdown, which is more informative than a single rank: a model can lead on text and restoration yet lose on identity preservation, which is the category that matters most for portraits and product shots. Second, the naming of OpenAI's variants is confusing. We have covered OpenAI's Flare and Sunburst image models, and the Microsoft headline and the leaderboard reports compare against different ones, which is why one says No. 2 and another says "above Flare."
How we got here: the MAI-Image timeline
The speed of Microsoft's releases is the real story. The company's news page lists, from newest back: MAI Image 2.7, MAI Image 2.6, MAI-Image-2.5-Pro, MAI-Image-2.5, MAI-Image-2-Efficient, MAI-Image-2 and MAI-Image-1.
- MAI-Image-2.5 (June 2, 2026): Microsoft said it ranked No. 2 on Arena's image-edit leaderboard, ahead of Nano Banana 2, and No. 3 for text-to-image. Foundry pricing was $5 per 1M text input tokens, $8 per 1M image input tokens and $47 per 1M image output tokens for the flagship, with a Flash tier at $1.75, $1.75 and $19.50. It also reached PowerPoint and OneDrive. See our post on the 2.5-Pro and Voice-2-Flash release.
- MAI-Image-2.6 and 2.6-Flash (September 2026): a speed-and-cost tier; Microsoft's post title for 2.6 says it launched at No. 2 on Arena. Covered in our 2.6 Flash post.
- MAI-Image-2.7-Preview (October 27, 2026): edit-only preview, claimed No. 2.
That is five image-model releases in under five months. Our 2.6 post predicted exactly this: any rank in such a post has "a shelf life measured in weeks." It also fits the wider pattern we described in Microsoft's hill-climbing approach to MAI models: ship quickly, measure on public boards, repeat. For the lab behind it, see MAI Base 2's pre-training claims.
Why a rank is not a verdict
Leaderboards like Artificial Analysis's and Arena's use human preference votes on paired outputs and convert them into Elo-style scores. That has real value: it reflects what people prefer when they see two edited images. But there are limits you should keep in mind.
- Preference is not correctness. A pleasing edit can still change a face, alter a logo or invent a detail. Microsoft's own 2.5 post warns that outputs "may produce plausible but inaccurate or misleading visual details" and should be reviewed in identity, legal, medical, financial or news workflows.
- Categories matter more than the headline. Four category wins of seven with a loss on identity preservation is a different profile from an across-the-board win.
- Preview models change. A preview can be retuned, rate-limited or renamed before general release, and board scores move as votes accumulate.
- Different boards, different ranks. The Artificial Analysis rank (No. 2 by the reports) and the Arena rank (No. 3 per Microsoft's post title) are not the same measurement.
- Cost and latency are missing. None of the reports we saw gave 2.7 pricing, a model ID or a speed figure.
What the category split tells you about use cases
The reported breakdown is worth reading as a map of strengths rather than a scoreboard. Scene and style edits, text or symbol edits and enhancement or restoration are the everyday jobs of marketing and e-commerce teams: swapping a background, fixing a typo on a banner, cleaning up a scanned photo. A model that leads there is attractive for content operations even if it is second overall.
Reasoning-based edits, where the model must understand the picture before changing it (for example, adding an object with the right perspective and shadows, which is how Microsoft described the 2.5 model), are the area most relevant to agents that edit images as one step in a longer workflow.
Identity-preserving edits and object-level edits are the opposite end. They are where a small drift is most visible, because a viewer notices when a face or a product shape changes. The reported gap there is the main reason to keep a human review step if you edit people or branded objects.
Availability and price: what exists today
For the models with published terms, Microsoft's 2.5 launch listed Foundry pricing of $5 per 1M text input tokens, $8 per 1M image input tokens and $47 per 1M image output tokens for the flagship, plus a Flash tier at lower rates, and said OpenRouter would also offer MAI-Image-2.5. We found no equivalent listing for 2.7 yet. Until one appears, budget with 2.6 or 2.5 numbers and treat 2.7 as a Playground experiment.
How to test it yourself in an afternoon
Do not route production traffic from a rank. A small, honest evaluation takes a few hours:
- Collect 30 to 50 real edits you actually need: remove an object, change on-image text, relight a product, restore an old photo, and several identity-sensitive portrait edits.
- Run the same prompts through MAI-Image-2.7-Preview in the MAI Playground and through your current model, such as GPT Image 2.5 or Google's Nano Banana 2.5.
- Score blind. Have two people pick the better output without knowing which model made it, and record failure types: changed faces, wrong text, ignored instructions.
- Check the cost per accepted edit, not per image. A cheaper model that needs three tries is not cheaper.
- Re-run in a month. If the model graduates to Foundry, repeat with the final model ID.
What this means for builders
- If you edit product photos or documents, the claimed leads in text edits and restoration are the ones to test first.
- If you edit people, the reported weakness on identity preservation is the reason to be careful. Test with consent-based examples only.
- If you are choosing a vendor, wait for Microsoft's 2.7 documentation, a Foundry listing and pricing. Until then the 2.6 and 2.5 models are the ones with published terms.
- If you follow the leaderboard race, treat OpenAI, Google, Meta, xAI and Microsoft as trading places every few weeks. The durable signal is your own evaluation set.
Open questions
- Is MAI-Image-2.7 going to Foundry, and at what price per image?
- Will the final release add text-to-image, and how will it rank there?
- How many votes sit behind the reported No. 2 placement and what are the confidence intervals?
- Does the model carry the same safety filters and provenance marking as earlier MAI-Image models?
We will update this post when Microsoft publishes the 2.7 documentation or the numbers can be checked on the live boards.
Related reading
- MAI-Image-2.6-Flash: Microsoft's faster, cheaper image model
- MAI-Image-2.5-Pro and MAI-Voice-2-Flash: what builders get
- Microsoft's MAI hill-climbing for Copilot and Excel
- OpenAI ChatGPT Images 2.5: Flare and Sunburst
- Google Nano Banana 2.5
- MAI Base 2: pre-training efficiency claim
Sources: Microsoft AI news and Microsoft AI, MAI-Image-2.5 launch. The rank and category figures were reported from Artificial Analysis and Mustafa Suleyman's posts on X; we could not open the underlying posts.
Details are accurate as of October 27, 2026. Leaderboard positions change frequently.