Microsoft's MAI-Image-2.5 Lands at No. 2 in Image Edit Arena
Microsoft AI officially released its in-house image generation and editing model "MAI-Image-2.5" around June 2, 2026, scoring 1401 and ranking No. 2 in the Single-Image-Edit category of the community-vote benchmark "Image Edit Arena," according to an announcement by Arena.ai's official account (@arena). The model reportedly beat Google's Nano Banana 2, xAI's Grok Imagine Image Quality, and OpenAI's ChatGPT-Image-Latest-High Fidelity by about 10 points each, advancing the Pareto frontier that represents the trade-off between quality and efficiency.
According to Microsoft AI's official account (@MicrosoftAI), MAI-Image-2.5 also placed No. 3 in the Text-to-Image Arena (score around 1254), marking an improvement of roughly 72 points over the previous-generation MAI-Image-2. Its Image Edit Arena score was 1401 (±8, with around 5,625 votes as of June 1, 2026), and its strength categories include text rendering, portraits, commercial motifs, photorealism, and prompt following. Alongside the release, a faster, lower-cost version called "MAI-Image-2.5-Flash" also became available.
Microsoft had previously used OpenAI's DALL·E-family models in products such as Bing Image Creator, but in recent years it has pushed its own first-party image models under the "MAI (Microsoft AI)" brand. Evolving from MAI-Image-1 to MAI-Image-2 and now MAI-Image-2.5, the lineup first entered the top contention zone in Text-to-Image in late May 2026 and gained higher rankings in editing capabilities in early June. Tech outlet The Decoder reported that MAI-Image-2.5 pulls even with Google's Nano Banana 2 on benchmarks and has improved substantially over its predecessor.
The focus of this advance is enterprise practicality. In creative workflows such as product photography, branding, and packaging, it leads competitors particularly in text rendering, instruction following, and commercial imagery. It can be tested instantly via MAI Playground, and supports API and enterprise use through Azure AI Foundry. Integration with PowerPoint is already complete, and rollout to OneDrive is underway. Pricing is per-token via Azure Foundry, listed at image output $47/M, text input $5/M, image input $8/M, with the Flash version offering image output at $19.50/M and text input at $1.75/M for lower cost and higher speed. Arena uses a blind comparison community-vote format, putting the model in direct competition with OpenAI (GPT Image family), Google (Nano Banana / Gemini family), and xAI (Grok).
On X, real users posted positive reports calling it "surprisingly good" and strong at iterative tasks such as room transformations, style-consistent edits, and fine detailing, while news of third-party integrations like fal.ai also circulated. At the same time, there were cautious voices noting that it "tends to drift in multi-step refinements," that "text rendering can be inferior to Grok Imagine," and pointing to a gap between hype and reality. Overall, developers and creators tended to praise it as "commercially oriented and practical" and welcomed Microsoft's progress on its own models, though some noted that in direct comparison with OpenAI and Google it has "not yet taken a clear lead."