Microsoft’s MAI Image 2.7 Ranks Above Google, Behind OpenAI

Where MAI Image 2.7 Ranks on Arena

Arena ranks image-editing models by asking people to vote between pairs of edited images. On its image-editing leaderboard on 28 October, OpenAI’s GPT-Image-2.5 Sunburst ranked first with a score of 1517. GPT-Image-2.5 Flare ranked second with 1482. Arena says both scores are preliminary.

MAI-Image-2.7-Preview ranked third with 1471. Its margin of error is plus or minus 10 points, based on 3,402 votes. Arena says its likely rank falls between second and fourth place. OpenAI’s GPT-Image-2 (Medium) ranked fourth with 1460, while xAI’s Grok Imagine Image 2.0 ranked fifth with 1439.

Microsoft’s earlier model, MAI-Image-2.6, ranked sixth with 1427. It had the same score as Google’s Gemini Nano Banana 2.1, which was Google’s highest-ranked model on the board. Based on our calculations, MAI Image 2.7 scored 46 points below Sunburst, 11 points below Flare, and 44 points above both MAI-Image-2.6 and Nano Banana 2.1.

What Mustafa Suleyman Posted on X

On 27 October, Mustafa Suleyman wrote on X that MAI-Image was now the world’s second-ranked image-editing model. He said it was once again beating Google, Meta, xAI, and GPT-Image-2.5-Flare. He also directed people to try the model on MAI Playground.

What Microsoft Says Has Changed

Microsoft AI says that “MAI-Image-2.7 improves on 2.6 across all 10 use cases and all 7 editing actions evaluated by Artificial Analysis.” The company says the model performs at the leading level in four areas: improving and restoring images, making edits that require reasoning, changing scenes and styles, and editing text or symbols inside images.

Microsoft says the model’s biggest advantage is in changing scenes and styles. In this area, it scores 24 Elo points more than the next-best model. Elo is a scoring system used to compare performance. Microsoft highlights three tasks: changing the style or lighting of a scene, changing text inside an image, and restoring existing images and other visual materials.

Microsoft says people can try the model on Arena, Artificial Analysis, and MAI Playground. When we checked Microsoft Foundry’s model catalogue for developers on 28 October, it listed MAI-Image-2.6 and MAI-Image-2.6-Flash, both marked as previews. It did not list MAI Image 2.7.

Other Releases That Day

On the same day, Microsoft AI released MAI Cyber 1.1 Flash. Microsoft says it delivers performance comparable to MAI Cyber 1 Flash while costing 18% less per scan in an internal test system called MDASH. MDASH is Microsoft’s system for testing and scanning security weaknesses.

Microsoft says that after customers run their first scan with MAI Cyber 1 Flash, nearly 9 out of 10 of their later scans use the MAI profile instead of the profile that uses only OpenAI models. Microsoft also says MAI Cyber now handles almost 50% of all scans performed through MDASH.

GitHub, which is owned by Microsoft, announced on 27 October that Copilot now supports server-to-server authentication for GitHub Apps installed by organizations. The feature is available as a public preview. Agents built with the Copilot SDK can send requests on behalf of an organization, and the organization that owns the app installation is charged for its use.

How This Fits with Earlier Reports

In June, we covered Microsoft Build and reported on Microsoft’s own AI models. In July, we reported that Microsoft was replacing OpenAI and Anthropic models with its own MAI models inside Excel and Outlook.

This month, we reported that Microsoft was building a model with 2.8 trillion parameters. A day ago, we also reported that Microsoft’s cybersecurity model had grown to account for 50% of scans. In image editing, Krea priced its Krea 3 image edits at $0.06 on 20 October. Krea 3 Ultra ranked ninth on the same Arena leaderboard.

What the Results Suggest

Our reading of Arena’s results puts MAI Image 2.7 in third place, behind OpenAI’s two newest models and ahead of every Google model listed on the board.

Inside MDASH, Microsoft says that once a customer has tried its MAI profile, nearly 9 out of 10 later scans use that profile rather than the profile that relies only on OpenAI.

GitHub’s change means that requests made by agents built with the Copilot SDK are charged to the organization that owns the app installation. This puts the costs under that organization’s own budget and spending controls.

Three Main Implications

FOR MICROSOFT Its image editor ranks third on Arena, and Microsoft says its cybersecurity model handles almost 50% of MDASH scans.

FOR OPENAI GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare are the only two models ranked above MAI Image 2.7 on Arena’s image-editing leaderboard. Both scores are preliminary.

FOR DEVELOPERS MAI Image 2.7 is not yet listed in Microsoft Foundry’s model catalogue. The catalogue lists MAI-Image-2.6 and MAI-Image-2.6-Flash as previews.

The Business Engineer View

This story relates to the Business Engineer framework called The Agentic Harness War.

The framework describes the change this way: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.” In simple terms, it argues that the software and tools surrounding an AI model are becoming an important source of value.

Microsoft’s MAI Cyber runs inside its own MDASH system, while GitHub now lets agents built with the Copilot SDK act on behalf of an organization. The framework was published in July 2026 and does not discuss these releases.

What Has Not Been Proven

Arena’s scores come from public votes, so they can change as more people vote. MAI Image 2.7 has received 3,402 votes, compared with tens of thousands to hundreds of thousands of votes for nearby models on the leaderboard. Its margin of error of plus or minus 10 points overlaps the scores of the models in second and fourth place.

Microsoft’s announcement gives no price for MAI Image 2.7 and no date for when developers will be able to access it. We did not contact Microsoft, OpenAI, or Arena for comment.

The Bottom Line

On 28 October, Microsoft AI’s MAI Image 2.7 ranked third on Arena’s image-editing leaderboard, with a score of 1471. It was behind two OpenAI models and scored 44 points more than Google’s highest-ranked model on the board. On 27 October, Microsoft also said its cybersecurity model now handles almost 50% of MDASH scans.