Microsoft's MAI-Image-2.5-Pro Beats OpenAI and Reve in Independent Editing Arena

Microsoft's newest image model just claimed the top spot on an independent image editing benchmark, less than a month after it entered preview. MAI-Image-2.5-Pro debuts at #1 on the Artificial Analysis Image Editing Leaderboard and #7 on the Text to Image side, giving Microsoft AI its first house model to lead an independent generative image arena.

MAI-Image-2.5-Pro currently leads the Artificial Analysis Image Editing Arena with an Elo score of 1271. The top image editing models by Elo rating are: MAI-Image-2.5-Pro (1271), Reve 2.1 (1263), GPT Image 2 high (1257), MAI-Image-2.5 (1256), and GPT Image 1.5 high (1250). Rankings are based on blind user votes in the Artificial Analysis Image Arena, so the ordering reflects human preference on side-by-side comparisons rather than a static reference set.

The quality tier of a three-model family

One of the core principles Microsoft pitches for Foundry is letting customers choose the right model for the right job. Some workloads need the highest possible quality, fidelity, and consistency for professional production. Others prioritize responsiveness and cost-efficiency for frequent, real-time responses. Pro sits firmly at the quality end of that curve, joining MAI-Image-2.5 and MAI-Image-2.5-Flash.

Under the hood, MAI-Image-2.5-Pro is a text-to-image and image-to-image editing model that uses a diffusion-based generative approach to progressively refine images. Microsoft says it excels at precise, surgical edits with consistency, enabling targeted object edits, layout changes, and text updates. The family also targets object removal, replacement, inpainting, and artifact cleanup while preserving composition, high-fidelity portraits with accurate facial structure and lighting, improved text rendering for labels, posters, and signage, and visual reasoning across scene structure and spatial positioning.

Pricing that puts quality at a premium

Pro is not cheap. Microsoft charges per token rather than per image, and the math works out to a meaningful step up from the rest of the family:

MAI-Image-2.5 Pro is available through Microsoft Foundry, with pricing starting at $5 per 1M tokens for text input and $106 per 1M tokens for image output, plus $8 per 1M image input tokens. That is more than twice the cost of the standard tier and roughly 5x the Flash variant, so the pitch only works if the fidelity gap justifies the bill for a given workload.

Why the debut matters beyond a leaderboard screenshot

The predecessor model already had strong footing. MAI-Image-2.5 ranked #2 in Text to Image behind only OpenAI's GPT Image 2, and its editing capabilities ranked #3, with performance comparable to Google's Nano Banana 2. Pro's arrival pushes Microsoft ahead of both OpenAI and Reve for editing tasks specifically, which is where enterprise creative workflows tend to spend the most compute.

The other signal is deployment surface. MAI-Image-2.5 is now the default model behind Bing Image Creator and specific PowerPoint and OneDrive paths, meaning Microsoft is already routing production traffic through its own family rather than paying a third party per generation. A benchmark-topping Pro tier gives the company a credible reason to keep steering high-end jobs internally too.

What builders get access to

• Preview availability on Microsoft Foundry across seven global-standard regions
• Interactive testing through the MAI Playground without wiring up the API
• Both text-to-image generation and controllable image-to-image editing from a single model
• Pay-as-you-go token pricing, with reserved throughput units available for predictable capacity

If your workflow is bulk generation or thumbnails, Flash at $20 per 1,000 images is almost certainly the right tier. If you are producing campaign hero imagery, doing precise object edits on customer photos, or rendering readable text inside a generated image, Pro is now the strongest independently-ranked option Microsoft offers, and the first MAI model to actually beat OpenAI's image stack on a public arena. The open question is whether the fidelity gap over MAI-Image-2.5 is large enough in practice to justify roughly doubling the per-image cost, which is a call that will come down to blind A/B testing on your own prompts rather than Elo points.